DEV Community: Prabhash Jha

How to Automate Your Work With AI: A Beginner's Guide

Prabhash Jha — Mon, 07 Sep 2026 15:02:38 +0000

Most advice about AI automation starts with the tools. Wrong end. Tools are the easy part. You can learn any of them in an afternoon. What actually decides whether automation saves you time or quietly wastes it is choosing what to automate. And almost nobody teaches that.

So this guide is the order I'd actually follow. Find the work worth automating. Understand the four kinds of tools and which one your task needs. Build the smallest possible version. Know in advance which mistakes turn a time-saver into a second job.

First, the uncomfortable maths

Every automation has a build cost and a maintenance cost. People remember the first. They forget the second.

Honest test before you automate anything:

Time saved per run × runs per month, versus build time + ongoing fixing.

Say a task takes 5 minutes and happens twice a month. You save 10 minutes. If building it takes three hours, you break even in eighteen months. And that's assuming nothing changes. Which it will.

A task that takes 10 minutes and happens daily saves around 5 hours a month. That's worth a day of building.

This one calculation kills roughly half of the automation ideas people get excited about. It should. The goal isn't to automate the most things. It's to stop doing the work that repeats.

Step 1: Find the work actually worth automating

Don't start by listing tasks you think are repetitive. Memory is a poor guide. You forget the small interruptions, which is usually where the time goes.

Try this instead. For one week, keep a rough log. Every time you do something that feels mechanical, write one line. What it was. Roughly how long. What triggered it. No tooling needed. A note on your phone is fine.

End of the week, look for tasks with all four of these:

1. It repeats on a predictable trigger. A form is submitted. An email arrives. A date passes. A file lands in a folder. Automation needs a starting gun. "Whenever I feel like it" cannot be automated.

2. The steps are the same every time. If you make a judgement call halfway through that changes what happens next, that judgement is the actual work. It either stays with you, or gets handed to a model deliberately. Not by accident.

3. The input is structured, or can be made structured. Data in a consistent shape automates well. This is where AI genuinely changed things. Messy input, an email in prose, a scanned invoice, a voice note, used to disqualify a task from automation. Not anymore.

4. A mistake is recoverable. Careful here. If an error sends a wrong invoice to a client, or emails the wrong list, the failure costs more than the work ever did.

Tasks that hit all four are your candidates. Rank them by the maths above. Start with the single highest one. Not three. One.

Step 2: Know which of the four tool types you need

Nearly every AI automation tool is one of four things. Picking the wrong category is the single most common reason a project stalls.

Type 1, chat assistants. ChatGPT, Claude, Gemini. You bring the input, they do the thinking, you take the output. No trigger. No schedule. A human is in the loop every time.
Right for: drafting, summarising, rewriting, analysis, thinking something through.
Wrong for: anything that needs to happen without you present.

Type 2, workflow builders. Zapier, Make, n8n. These connect apps. When this happens in one place, do that in another. Most now include AI steps, so a model can sit in the middle of the chain.
Right for: the actual repetitive work. Moving, routing, formatting, notifying.
Wrong for: tasks needing genuine judgement at every step.

Type 3, built-in AI inside tools you already pay for. Your spreadsheet, your CRM, your email, your docs. All of them ship AI features now.
Right for: the first thing to check, always. Free, no integration, nothing new to maintain.
Wrong for: anything crossing between systems.

Type 4, agents. Tools that take a goal and decide their own steps. Genuinely powerful. Genuinely immature.
Right for: exploratory work where the path isn't known.
Wrong for: anything where a wrong step is expensive. Which is most business processes today.

Most people reach for Type 4 because it's the most exciting, when the answer is Type 3 followed by Type 2.

Step 3: Build the smallest version that works

The instinct is to design the whole system. Resist it. Build the smallest thing that removes some of the work, run it for a week, then extend.

Concretely:

Do it manually once and write down every step. Including the ones you don't notice. The copy, the paste, the rename, the check. Automation fails on the steps you forgot were steps.
Automate only the middle. Leave the trigger manual (you press a button) and the output manual (you review before it sends). This is the version that catches your logic errors while they're still cheap.
Run it alongside the manual process for a week. Compare outputs. You're looking for the cases you didn't anticipate. And there will be some.
Automate the trigger. Now it runs without you starting it.
Automate the output last, and only if step 3 was boring. If you found surprises, keep a human check.

That order feels slow. In practice it's dramatically faster than debugging a fully automated chain that's been quietly producing wrong output for three weeks.

Where AI actually helps, versus where plain automation was always enough

This distinction saves a lot of money and complexity, and it's rarely made.

Plain automation is enough when the rule is clear. "When a form is submitted, add a row to the sheet and send a notification" needs no AI. It's a rule. Adding a model makes it slower, more expensive, and less predictable.

AI earns its place when the input is unstructured or the output needs language. Reading a paragraph and pulling out the three facts you need. Sorting incoming messages by what they're actually about. Turning bullet points into a first draft. Summarising a long thread. Categorising feedback that arrives as free text.

Good rule: if you could write the rule in a sentence, don't use a model. If explaining it requires "well, it depends what they mean," that's where a model belongs.

The five mistakes that make automation cost more than it saves

1. Automating a bad process. Automation makes a process faster, not better. If the underlying workflow is convoluted, you've built a machine for producing the wrong thing efficiently. Simplify first. Then automate.

2. No failure notification. Worst kind of automation is one that breaks silently. If a workflow stops running and nobody notices for a month, you've lost a month of that work with no record. Every automation needs a "tell me when this fails" step. Honestly, this is the single most-skipped thing. And a failure step is only half of it, because the nastiest version of this bug never fails at all: the workflow that runs green and does nothing reports success on every run while producing no effect. Which is why the job should report a count rather than a status.

3. Over-trusting the model's output. Language models produce plausible text. Not the same as correct text. Anything factual, names, numbers, dates, calculations, needs either verification or a step that doesn't rely on a model. Never let an unverified model output touch a customer.

4. Building on a tool you don't understand the pricing of. Workflow tools generally charge per successful action, and a chain that loops can consume a month's quota in an afternoon. Understand what one run costs before you schedule it hourly.

5. Nobody else can fix it. Build automation for a team, and if only you understand it, you've created a dependency, not leverage. Write down what it does. What triggers it. And how to turn it off. Especially how to turn it off.

A worked example

Common one. Incoming enquiries arrive by email and need logging, categorising and responding to.

Manual, that's maybe 6 minutes each. At 5 a day, roughly 10 hours a month. Clearly worth automating. Comfortably above the 5-hour-a-month floor the honest should-you-automate-this arithmetic sets, and with a maintenance term a well-designed API-backed version keeps small.

Here's the build, in the order above:

Trigger: a new email arrives at a specific address.
AI step: extract the sender's name, company, what they're asking about, and how urgent it sounds. This is the part that needed a human before. The email is prose, not a form.
Rule step: write those fields to a sheet or CRM. No AI needed. It's a rule.
Branch: if the enquiry matches a known category, draft a reply from a template. If it doesn't, flag it for a human.
Human gate: the draft goes to you, not the sender. You read and send.
Failure alert: if the workflow errors, notify.

Notice where the model sits. Exactly one step. Doing the one thing rules can't, which is reading unstructured language. Everything else is deterministic. That's what a well-built automation looks like. The opposite of handing the whole task to an agent and hoping.

Note too that the human gate stays. You could remove it once you've watched the drafts for a month. You probably won't want to, and that's a legitimate choice. The 6 minutes is now 30 seconds either way.

How to actually learn this

You don't learn automation by reading about it. The path that works:

Pick one task from your log. The highest on the maths. Not the most interesting.
Build it badly. It will be inelegant. It will still save you time.
Live with it for a month. You'll discover the edge cases. And you'll learn more from fixing them than from any tutorial.
Then build the second one. It'll take a quarter of the time.

Most people stall because they try to learn the tool comprehensively before building anything. Automation is a skill you acquire by having built things. Same as marketing.

FAQs

What is AI workflow automation?

It's connecting your apps so a sequence runs without you, with an AI model handling the steps that require interpreting language or unstructured input. The automation part moves and routes data. The AI part reads, classifies, extracts or writes. Most useful workflows are mostly rules with a model doing one specific job in the middle.

Which tasks should I automate first?

Ones that repeat on a predictable trigger, follow the same steps every time, and where a mistake is recoverable. Rank candidates by time saved per run multiplied by runs per month, against the hours it'll take to build and maintain. Daily 10-minute tasks are almost always worth it. Twice-monthly 5-minute tasks almost never are.

Do I need to know how to code to automate work with AI?

No. Workflow builders are visual. You connect blocks. Knowing a little scripting widens what you can do, but the constraint for most people isn't coding ability. It's process clarity. If you can't describe the steps precisely in writing, no tool will help.

What's the difference between AI automation and an AI agent?

Automation follows a path you defined. An agent decides its own path toward a goal. Automation is predictable and debuggable, which is why it suits business processes. Agents are more flexible and less predictable, which suits exploration. Start with automation.

How much does it cost to automate work with AI?

Workflow tools typically charge per task or operation, and model usage is billed by volume of text processed. A single well-scoped workflow running a few hundred times a month usually costs less than an hour of your time. The risk isn't the price per run. It's a looping workflow burning through quota unnoticed. Which is why understanding pricing before scheduling matters.

Will AI automation replace my job?

It replaces tasks, not roles. The parts most exposed are the mechanical, repeatable ones. Which are also the parts you least enjoy. The realistic risk isn't being replaced by AI. It's being outpaced by someone doing the same role with the mechanical parts removed.

Key takeaways

Choosing what to automate matters far more than which tool you pick.
Run the maths first: time saved per run × frequency, against build plus maintenance.
Check the AI already inside tools you pay for before adding anything new.
Use a model only where rules can't work, unstructured input or language output.
Build the middle first, keep humans on the trigger and the output, automate those last.
Every automation needs a failure alert. Silent breakage is the expensive failure mode.

How to Actually Get Value From Marketing Events and Conferences

Prabhash Jha — Mon, 07 Sep 2026 15:02:37 +0000

You paid for the ticket. Sat through nine talks. Filled four pages of notes. Came home with a branded notebook and a vague sense you should be doing more video. Three weeks later nothing in your business has changed. So you decide conferences are a waste of money.

They usually are. But not for the reason you think. The mistake is going for the talks.

A conference isn't a training course. Almost everything said on stage is a compressed, sanitised version of something already published somewhere, minus the numbers that would make it useful. If education were the point, a day of reading would beat a day of sitting. The stage isn't a curriculum. It's a filter. It tells you which people in the building are worth twenty minutes of your time.

The event isn't the value. It's the doorway. What matters is the specific people you meet and what you do in the two weeks after. And the pattern is that the useful ones tend to be casual acquaintances rather than existing close contacts, which is the finding of a twenty-million-person LinkedIn study on the strength of weak ties. Job mobility rises fastest with acquaintances two hops out, not with your inner circle, because those are the people who see opportunities you haven't already heard about. Here's how I get real value from marketing events, and how to work out honestly whether a given one is worth attending at all.

Why most people waste marketing events

They show up with no goal. They sit through talks passively. They network randomly. They never follow up.

Each of those has a mechanism worth naming. Passive listening feels productive because it produces artefacts. Notes, photos, a LinkedIn post about how insightful the panel was. None of it survives contact with Monday. "No goal" is usually a polite way of saying "no list": you didn't decide before you arrived which twelve humans you were there to meet, so you met whoever was standing near the coffee. And the follow-up failure isn't laziness. It's capacity. People come back to a full inbox and two days of missed work, and the fortnight where the relationship was still warm gets eaten by catch-up.

Fix those four and an average event outperforms a great one you attended badly. That's not a small claim. It's the whole game. You don't need the best conference in your industry. You need a plan that survives the ride home.

What a conference actually gives you that the internet doesn't

Be specific about the goods. It changes what you do on the day.

Proximity to people who ignore cold email. A head of growth who never answers LinkedIn will happily talk for fifteen minutes in a lunch queue, because they're wearing a badge and standing still. That's the entire product you bought.

Price and vendor truth. What people genuinely pay for a tool. Which platform rep actually returns calls. Which vendor's onboarding is a nightmare. This circulates in corridors and never in blog posts. Two honest conversations about spend and rates can save you more than the ticket.

Calibration. You find out whether your results are good, average or embarrassing for your category. Most people work in a vacuum and either overrate or underrate themselves by a wide margin.

Speed of trust. Twenty minutes face to face does what twenty emails cannot. Deals that would take six months of nurturing compress into weeks because the person has a face, a handshake and a memory attached to your name.

None of that happens on stage. It happens in the gaps. In the queue. In the smoking corner. In the awkward five minutes before a session starts. The stage is the excuse for everyone being in the same building.

How to prepare: the two weeks before

Set one clear goal. Three useful conversations. One partner. One specific thing to learn. One goal beats ten vague ones. Write it as an outcome, not a theme. "Learn about AI in marketing" isn't a goal. "Leave with two agency partners who handle CRO and don't do media buying" is.

Build the list. The speaker roster is public weeks ahead. Sponsors publish which of their team is attending. People announce themselves on LinkedIn with the event hashtag. The event app usually opens three to seven days early with a searchable attendee directory and in-app messaging. That directory is the most underused asset in the whole event. Get to fifteen names, ranked A, B, C. Fifteen is the ceiling of what two days can realistically hold.

Research one specific thing per name. A recent post. A launch. A job advert they've put out. A change of role. That single detail is your opening line and the reason they'll remember you an hour later. "Saw you're hiring two performance people, are you bringing media buying in-house?" beats "so what do you do?" every time.

Prepare your one-liner. A clear, human answer to "what do you do?" that invites a follow-up question. The test is simple. Does the other person say "how?" or do they say "nice"? If it ends the conversation, rewrite it. Who you help, the problem you remove, one concrete detail.

Clear the fortnight after. This is the preparation step nobody does. Block three hours in the week following the event before you book the ticket. If you can't protect that time, you're buying a doorway you won't walk through.

Boring logistics that decide your day. Power bank, because the phone is your notes, your calendar and your CRM. Cards, because in India they still work, and fumbling for a QR code in a loud hall loses the moment. Something identifiable to wear, so "the guy in the green jacket" is you when someone describes you to a colleague.

Pre-event outreach: how to book meetings before you arrive

The people worth meeting have their two days spoken for by the time they land. Yours should be too.

Send your first round seven to ten days out. One message per name. No template smell. Three sentences: you'll both be at the event, the specific reason you want to talk to them, and a concrete ask. Fifteen minutes, day one, 11:30, at the coffee counter near registration. Name a place and a time. "Let's find each other there" is how meetings don't happen.

Don't pitch in the request. The ask is for time, not for business. The moment it reads as a sales approach your reply rate collapses. Send a one-line reminder the day before with your phone number and what you'll be wearing.

Expect four to six yeses out of fifteen asks, and expect one or two to fall through on the day. That's a normal, good outcome. It means your calendar has structure and the gaps are free for the accidental conversations, which are often the best ones.

What to actually do during the sessions

Sessions aren't useless. They're just not what you think they're for.

Pick by speaker, not by title. Titles are written by the marketing team. The speaker is the actual variable. Sit on an aisle near the front, and give yourself explicit permission to walk out after eight minutes if it's thin. Nobody minds and you get the corridor to yourself.

Use the Q&A as your business card. Stand up. Give your name and company in one clause. Ask a question that is specific enough that only a practitioner could have asked it. Two or three people will find you afterwards. It's the cheapest inbound available at the entire event and it costs you thirty seconds of discomfort.

Take notes on people, not just talks. One line about each person and a reason to follow up. Write it into your phone within sixty seconds of walking away. Not at the end of the day, when six faces have merged into one. Name, company, the thing they're wrestling with, anything you promised to send.

Log the audience's questions. What people ask from the floor is unfiltered demand. The phrasing they use. The thing they're stuck on. The objection they can't get past. That's a free keyword research session and a content calendar. See SEO basics: how to get found on Google without paying for ads for what to do with it.

How to work the hallways without being that person

Prioritise hallways over stages. The best conversations happen between sessions, not during them. Talks you can usually watch later.

Give before you ask. Make an introduction. Share a useful idea. Solve a small problem. Generosity is the fastest way to be remembered. And the only reliable way to be remembered by someone more senior than you.

Quality over quantity. Five real conversations beat fifty forgettable ones. A stack of fifty cards is a record of fifty people who don't remember you.

Practical geography. The registration desk between 8:30 and 9:30, when people are alone and unhurried. The coffee queue, where standing still is socially mandated. Lunch tables, where you sit with strangers rather than colleagues. The last ninety minutes of day one, when everyone has relaxed and the pretence is gone. The worst venue is the sponsored party after 9pm. High volume, low recall, no notes taken.

Two questions do most of the work. "What's actually working for you right now?" and "what's broken?" Both are easy to answer, neither is a pitch, and the second one is where every real opportunity lives.

Never sell in a first conversation. Ask about the problem, and if there's a genuine fit, ask permission to send something specific. "Can I send you the way we structure that?" gets a yes. A pitch gets a polite exit. And learn to leave a conversation cleanly. "I'll let you get to your next session, I'll message you tonight" so you can have ten of them instead of two. Same rules as networking generally, just compressed into two days.

Which marketing events are worth the money?

Not all events do the same job. Costs below are rough Indian ranges including travel where relevant. Check current pricing, but the shape holds.

Event type	Typical loaded cost	Who's actually in the room	Best for	Honest expectation
Large national marketing conference	₹40,000–₹1,00,000	Brand-side marketers, agencies, platform reps, a lot of juniors	Broad calibration, platform relationships, one or two strong leads	Low hit rate, high ceiling, needs a strict list to be worth it
Vertical or industry summit (BFSI, D2C, gaming, SaaS)	₹50,000–₹1,50,000	Decision-makers in one category, fewer people overall	Category credibility, buyers who already have your problem	The best ratio of the lot if the vertical is genuinely yours
Affiliate / performance meet-up	₹80,000–₹2,50,000 with travel	Networks, advertisers, media buyers, tool vendors	Deal terms, payout and tracking realities, partner supply	Deal-driven. Go with commercial terms ready or don't go
Awards night	₹15,000–₹60,000 a seat	Peers, press, a lot of drinking	Recruiting, visibility, peer relationships	Almost never produces revenue. Treat it as brand and hiring spend
Local meetup or community event	₹0–₹2,000	Practitioners within an hour of you	Repeat contact, building a local reputation	Compounds over months. Single visits do nothing
Virtual conference	₹0–₹10,000	Whoever registered. Most never watch	The recordings and the attendee list	Useful for research, useless for relationships

Read the row that matters for you and note the pattern. The cheaper the room, the more repetition it needs. The more expensive, the more preparation it demands.

The follow-up window that decides everything

Follow up within 48 hours, while you're still fresh in their memory. A short, specific, personal message. Not a copy-paste. Weekend event? Sunday night is fine. Wednesday is too late, because by then four other people have messaged them and none of you are distinguishable.

Reference the actual conversation so they remember you, and suggest one clear next step. Four lines is enough. Where you met and what you talked about. The thing you promised (attached, now). One sentence of genuinely useful substance. One specific ask with a time. Any message that could have been sent to sixty people gets treated like it was.

Channel: email if you took an email. LinkedIn if you didn't. WhatsApp only if they handed you the number themselves. Using a number scraped from a badge photo reads as a trespass.

Then the second touch, around day ten to fourteen. Not "just bumping this up." Send something new. A relevant article. An introduction you promised. An answer to the thing they were stuck on. After that, move them to a quarterly low-effort touch. This is exactly what an owned list is for. See how to build an email list from zero.

Act on what you learned. Turn the best idea into one small experiment this week, before the motivation fades. One, not five. Five is how nothing gets done.

And keep one sheet. Name, company, where you met, what you discussed, what you promised, next step, date. If a person isn't in that sheet before you land, that person doesn't exist.

How to measure conference ROI honestly

Run the numbers on a hypothetical flagship conference, using your own figures when you do this for real.

Cash cost: ticket ₹25,000 + return flight ₹11,000 + two hotel nights at ₹6,500 = ₹13,000 + food and local transport ₹4,000 = ₹53,000.

Time cost: two event days plus one travel-and-prep day. If a working day of yours is worth ₹12,000 in output you'd otherwise have produced, that's ₹36,000.

Loaded cost: ₹89,000.

Now the pipeline. Say your follow-ups convert at: 60% reply, half of repliers take a call, half of those calls surface a real opportunity, and you close 40% of opportunities. Multiply: 0.6 × 0.5 × 0.5 × 0.4 = 0.06 clients per genuine follow-up. Roughly one client for every seventeen. If a client contributes ₹1,50,000 in gross profit over the relationship, each genuine follow-up carries about ₹9,000 of expected value.

That single number reframes the whole event:

Genuine follow-ups sent	Expected clients	Expected gross profit	Against ₹89,000 cost
6	0.36	₹54,000	−₹35,000
10	0.60	₹90,000	+₹1,000
15	0.90	₹1,35,000	+₹46,000

You break even at ten follow-ups. Almost nobody gets there, because the follow-up happens after the excitement has worn off and the inbox has refilled. That's the entire difference between "conferences don't work" and "conferences work." Not the venue. Not the speakers. Not the ticket price.

Two more things the table shows. First, the cost side is a lever too. Attend a local event with no flight or hotel and the loaded cost drops to about ₹53,000, where break-even arrives at six follow-ups instead of ten. Second, if you can't plausibly produce ten real conversations in two days, either the room is wrong for you or you needed a better list.

Keep a second ledger for value that isn't revenue. A supplier rate you renegotiated. A hire you found. A bad tool you avoided buying. It's real. It also isn't infinite. If the second ledger is the only thing keeping an event alive on paper two years running, you're rationalising. Tag the source on any deal that traces back, use a six-to-twelve month window, and don't credit the conference for a deal that was already in your pipeline. The general discipline is in the sheet that tells you whether your marketing makes money, and the metric definitions are in marketing metrics explained.

Honestly, most events I've been to have failed on that second ledger too. The ones that worked, worked because I already had a plan.

When to skip an event

Skip it when you can't name six people you want to meet after twenty minutes of looking at the attendee list. Skip it when the room is mostly people selling what you sell. Agencies pitching agencies is a closed loop. Skip it when the two weeks after are already full, because you'll buy the doorway and not walk through it. Skip it when it's the same crowd as last year and last year produced nothing. Skip an awards night that costs a month of marketing budget unless you're actively hiring or raising.

There are cheaper substitutes for the same goods. Go to the host city without a ticket. The side events, hotel lobbies and dinners are where half the business happens and they're free. Host a dinner for eight of the right people for less than one flagship ticket. Apply to speak instead of attending. A speaker gets inbound rather than chasing it. Or run a small local meetup, which builds the same relationships slowly at almost no cost.

If you're exhibiting or sponsoring

Different game, same principle. Stand traffic is mostly people collecting merchandise, and the badge-scan count your sales team celebrates is close to meaningless. What a stand actually gives you is a fixed location for meetings you booked beforehand, plus credibility for the people who were already going to talk to you.

The budget rule. If you can't fund the follow-up, actual named people with actual blocked time to work the list within a week, don't buy the stand. A stand with no follow-up capacity is the most expensive way to hand out pens ever invented.

FAQs

Are marketing conferences worth it?

Yes, if you go with a goal and follow up. The ticket pays for itself through one good relationship or idea. Without a plan, it's an expensive day out. Run the arithmetic before you book. Loaded cost divided by the gross profit of one client tells you how many genuine follow-ups you need. If that number is more conversations than the event can hold, skip it.

How do I network at events if I'm introverted?

You don't need to work the room. Have a few genuine one-on-one conversations, lead with curiosity and helpfulness, and follow up in writing, where introverts often shine. Depth beats volume. Book two or three meetings in advance so you arrive with structure rather than having to improvise, and take deliberate breaks. Five real conversations beat fifty forgettable ones.

How soon should I follow up after a conference?

Within 48 hours, while the conversation is still specific in their memory. After roughly four days you're competing with everyone else who met them, and you all sound the same. Send the promised thing immediately rather than saving it for a "proper" email later. The polished version you send next week is worth less than the rough one you send tonight.

Do business cards still work, or should I use a QR code?

Both, and carry cards. In a loud hall with weak Wi-Fi, a card is faster and it doesn't die with your battery. The point isn't the card anyway. It's what you write on the back of it in the ninety seconds after the conversation. A card with a note is a record. A scanned profile with no context is a stranger.

How do I convince my company to pay for a conference ticket?

Ask for it as a pipeline activity, not training. Show the list of named people you intend to meet, the loaded cost, and the number of follow-ups you'd need for it to pay back. Offer a written debrief within a week. Contacts made, next steps booked, one experiment to run. Managers refuse "I'd like to learn." They approve a plan with a number attached.

Are virtual conferences worth attending?

For relationships, rarely. Nobody remembers a face in a grid. For research they're excellent and often free. You get the recordings, the chat, the questions people ask and, most usefully, the attendee list. Treat a virtual event as a source of names and demand signals to work through afterwards, then take those relationships to a real room.

Key takeaways

Treating a conference as education is why people conclude conferences are a waste. The stage is a filter for who to meet, not a curriculum.
The event itself isn't the value. It's the doorway. What matters is the specific people and what you do in the fortnight afterwards.
Preparation is a named list of about fifteen people plus pre-booked meetings with a stated place and time, not a highlighted agenda.
Give before you ask and take notes on people rather than talks. Five real conversations beat fifty forgettable ones.
Follow up within 48 hours referencing the actual conversation. On the worked example above each genuine follow-up carries roughly ₹9,000 of expected value, so how many you send matters more than which event you picked.
Block the follow-up time before you buy the ticket, and skip any event where you can't name six people worth meeting.

How Much Runway Does a Services Business Actually Need?

Prabhash Jha — Sun, 06 Sep 2026 12:52:13 +0000

Every runway calculator you can find asks for two inputs: cash in the bank and monthly burn. It divides one by the other and tells you how many months you have. That model was built for venture-funded product businesses, where the assumptions behind it are all true, revenue is small and predictable, costs are largely engineering, and the thing you are counting down to is the next funding round.

Almost none of that describes a services business, and the model fails in a specific and dangerous way when you apply it anyway. It tells you that you are fine.

Search the question and you will find plenty written about client concentration, but nearly all of it treats concentration as a valuation problem, how a buyer discounts your business at exit, what a lender thinks of your credit file. That is a real concern for a firm being sold. It is not the concern of a founder trying to work out how much cash to keep in the bank this year, and the two need completely different arithmetic.

The number a services business needs is not months of burn. It is the cash required to survive losing your largest client, for as long as it takes to replace them, while paying people you cannot legally or practically shed on that timeline. That is a different calculation with a different answer, and it is usually a larger one.

Why months-of-burn is the wrong model here

The burn model assumes costs are flexible and revenue is stable. In a services business both assumptions are inverted: costs are close to fixed and revenue can be cancelled with notice.

Three structural differences do the damage.

Your costs are people, and people are semi-fixed. A product business under pressure can cut cloud spend, pause marketing, and defer hiring, and its cost base genuinely moves within a month. Your cost base is salaries. Salaries do not fall when revenue does. They fall when you make people redundant, which involves notice periods, statutory dues, and a real cost to execute, and which permanently reduces your capacity to deliver the work you still have. Cutting costs in a services business is not a dial. It is a one-way door with a fee attached.

Your revenue is cancellable on notice, in blocks. A subscription business loses customers one at a time, in a curve you can forecast. You lose them in lumps, on thirty or sixty days' notice, and one lump can be a quarter of the business. Nothing about the previous six months predicts it, because the decision is made in a meeting you are not in.

Recognised revenue is not cash. The burn model quietly assumes what you invoice is what you bank. In services the gap between the two is the whole problem, and it widens exactly when things get difficult, clients under pressure pay later, and a client on their way out pays latest of all. If that distinction is not already sharp for you, cash flow vs profit is the foundation this post is built on.

Put together, these mean a services business under stress has costs that are slow and expensive to reduce, revenue that can disappear in a single conversation, and collections that stretch at the worst moment. The burn model sees none of this. It sees a healthy business with eight months of cover.

The question that actually sets the number

Ask what happens if your largest client leaves tomorrow, and how long it takes to replace that revenue, not how long you can survive with no revenue at all.

The all-revenue-stops scenario is the one people model, and it is the wrong one. It is both too pessimistic (all your clients will not leave in the same month) and, more importantly, useless, it produces a number so large you dismiss it, which means you end up planning against nothing at all.

The realistic catastrophic scenario for a services business is narrow and specific: your largest client gives notice. Everything else keeps running. That is the event to size against, because it is the one that actually happens, and it has three parameters you can measure today.

One: your fixed cost floor. Not your current monthly spend, the floor you can realistically get to within one quarter without breaking your ability to serve the clients you still have. Salaries of people you will keep, rent, software, compliance, the minimum of everything else. Most founders have never separated their spend into "floor" and "above floor," and the exercise is uncomfortable, because the honest floor is usually much closer to the current number than expected.

Two: replacement time. How long from a client giving notice to a replacement of similar size paying you, which is your sales cycle, plus contracting, plus onboarding, plus their first payment terms. That final leg is the one everybody forgets. A ninety-day sales cycle with sixty-day payment terms means the money arrives five months after you start looking, and the search does not start on the day of the notice; it starts a couple of weeks later once you have accepted that it is real.

Three: your concentration. What proportion of monthly revenue your largest client represents, and the same for your top three. Do this on gross revenue and again on margin contribution, because they are frequently different clients, the largest by fee is not always the largest by profit, which is the entire subject of which client is actually profitable when everyone works on everything. The margin figure is the one that matters here.

The arithmetic, worked through

Take a hypothetical services business, purely to show the mechanics. Monthly revenue of ₹20,00,000. Fixed monthly costs of ₹16,00,000, of which ₹13,00,000 is the floor, the team, rent and systems that have to survive for the remaining clients to be served. The largest client is ₹6,00,000 a month, or 30% of revenue. Notice period is sixty days. The sales cycle for an account of that size runs about ninety days, and new clients pay on thirty-day terms.

The burn model says: ₹4,00,000 a month of surplus, comfortably profitable, and with ₹20,00,000 in the bank you have "infinite runway." No alarm anywhere.

Now run the actual event.

Month	What happens	Monthly cash effect
0	Client gives notice; still paying, still being served	+₹4,00,000
1–2	Notice period runs out; search starts around week 3	+₹4,00,000, then ₹0
3	Revenue now ₹14,00,000 against a ₹16,00,000 cost base	−₹2,00,000
4–5	Cost cutting begins; you reach the ₹13,00,000 floor by month 5	−₹2,00,000, then +₹1,00,000
5	New client signs after a ~90-day cycle from month 2	+₹1,00,000
6	Onboarding; first invoice raised	+₹1,00,000
7	First payment lands on 30-day terms	+₹7,00,000

The cash trough runs roughly months 3 to 6, and the deficit accumulated across it is on the order of ₹4,00,000 to ₹5,00,000, modest, and easily covered by the ₹20,00,000 balance. This business is genuinely fine.

Change one input. Make the largest client ₹10,00,000 a month, 50% of revenue, which is not unusual for a firm that grew on the back of one relationship. Now revenue drops to ₹10,00,000 against a ₹13,00,000 floor, the monthly deficit through the trough is ₹3,00,000 rather than ₹2,00,000, the trough is longer because replacing a ₹10,00,000 account takes longer than replacing a ₹6,00,000 one, and, the part that does the real damage, you cannot cut to the floor, because the floor was calculated to serve a client who has left, and cutting further means you cannot win the replacement. The same balance now looks thin, and it got there by changing one variable that the burn model does not ask about.

That is the point of the exercise. Two businesses with identical revenue, identical costs, identical cash and identical burn have entirely different risk, and the only variable that separates them is concentration.

Concentration-adjusted runway, as a rule you can apply

Size your buffer as: months to replace your largest client, times the monthly deficit you would run after cutting to your floor, plus one quarter of the floor as a margin for the things that go wrong at the same time.

Written out:

Buffer = (replacement months × post-loss monthly deficit) + (0.25 × annual fixed floor ÷ 4)

The three components, and why each is there:

Replacement months. Sales cycle, plus contracting, plus onboarding, plus first payment terms, plus two weeks of denial. Use your actual longest recent cycle, not your average, you are sizing against a bad case, and averages are made of good cases too.

Post-loss monthly deficit. Revenue after the loss, minus your realistic floor. If that is positive, you do not have a runway problem for this client, and you can stop. Run the calculation again for your top three combined, because a downturn that takes one client often takes two.

The quarter of floor. This is not padding. It is there because the events correlate: the quarter you lose a large client is disproportionately the quarter another one stretches payment from thirty days to sixty, a receivable goes bad, and something breaks that needs money. If a payment problem is already live, the escalation sequence is the client has stopped paying, and the reason this term exists is that the two situations turn up together far more often than independence would suggest.

Two adjustments for how the business is actually structured:

Project work needs more than retainer work. A retainer book gives you notice periods and a predictable pipeline. A project book can simply end, with everything finishing in the same quarter and no notice at all. If most of your revenue is project-based, treat your replacement time as longer and your notice period as zero.
A senior-heavy team needs more than a junior-heavy one. Senior people are harder to re-hire, so cutting them is more costly to reverse, which means you will, correctly, resist cutting them, which means your realistic floor is higher than the one on paper. Be honest about this when you calculate the floor rather than discovering it in the moment.

Why "three to six months of expenses" is wrong in both directions

The standard advice is wrong for a low-concentration business and dangerously wrong for a high-concentration one, because it is indexed to expenses rather than to exposure.

For a firm with twenty clients and no single account above 8% of revenue, six months of expenses is a large amount of dead capital. No plausible single event takes out enough revenue to need it. That business can hold considerably less and put the difference into hiring, or capability, or simply take it out, the risk it is insuring against does not exist in the shape the rule assumes.

For a firm where one client is half the revenue, six months of expenses may not be enough, because the relevant clock is not "how long until I run out of money" but "how long until I have replaced that revenue," and those are unrelated quantities. If your replacement cycle genuinely runs eight months, six months of cover means the business fails on month seven with a signed contract in hand and nothing to bridge to it. That is a distinctly ugly way to go, and it is what the expenses-based rule quietly permits.

The rule is popular because it is easy and because it works passably for personal finance, where the model does fit, an individual's income really does stop entirely, and the expenses really are the number. That version is a genuinely good rule and is covered properly in how to build an emergency fund. Applying it unchanged to a business with lumpy, cancellable, concentrated revenue is where it breaks.

The Indian specifics that change the number

If you bill in India, several structural cash effects sit between "the client agreed to pay" and "you have the money," and they all push the buffer up.

GST is paid on the invoice, not on the payment. You remit output GST for the month you invoiced, regardless of whether the client has paid. That is the statute rather than a quirk of practice: section 13(2) of the CGST Act fixes the time of supply for services at the earliest of the date the invoice is issued or the date payment is received, so raising the invoice is itself the taxable event. On a large invoice with long terms, that is a genuine outflow funded by you, in advance, on money you have not received. TDS is deducted before you see it. Your client withholds tax at source, so the cash landing in your account is below the invoice value, and you recover the difference much later against your own liability. Both of these are covered properly in GST and TDS both hit your cash before the client does, and the practical consequence for this calculation is that your cash revenue in any month is meaningfully below your invoiced revenue while your costs are at full value.

Statutory dues on exit are real money. If part of your response to losing a client is reducing headcount, notice pay, gratuity where applicable and leave encashment all land in the same quarter as the revenue loss. The cost of shrinking is front-loaded and the saving is back-loaded. Model the cut as an outflow first.

If you bill abroad, the currency moves against you at inconvenient times. Foreign receivables convert at whatever the rate is on the day, not the day you quoted, and that exposure sits on top of everything else, billing abroad, spending at home covers how to think about it.

If the number comes out uncomfortably large

Reduce the exposure rather than trying to save your way to the buffer, because the buffer is a symptom and the concentration is the disease.

Most founders who do this calculation for the first time get a number they cannot fund out of current profit in any reasonable period. That is useful information, not a failure. The buffer is one of four levers and usually the slowest:

Shorten replacement time. The largest single input in the formula. A pipeline that exists before you need it, a handful of live conversations you maintain even when full, cuts months off the worst case, and costs almost nothing but discipline. Most services businesses do zero business development while busy, which is precisely what makes the replacement clock long.

Reduce the concentration directly. Not by refusing good work, but by deliberately weighting new business toward accounts that reduce the ratio, and by being clear-eyed that a very large client is a commercial risk as well as a commercial win. Growing the denominator works as well as shrinking the numerator.

Improve the terms. Shorter payment terms, longer notice periods, and part-advance billing all reduce the depth of the trough without needing a rupee more in the bank. If you are Udyam-registered as a micro or small enterprise, part of this lever is already law and most founders never use it: section 15 of the MSMED Act, 2006 provides that a period agreed in writing between supplier and buyer may in no case exceed forty-five days from the day of acceptance or deemed acceptance of the goods or services, and section 16 makes a late payer liable for compound interest, with monthly rests, at three times the bank rate notified by the RBI. You will rarely invoke it against a client you want to keep. But it sets the ceiling any negotiation about terms starts from, and knowing that a 90-day term is unenforceable against you changes how the conversation goes. Notice period is the most under-negotiated term in most services contracts and it is worth real money in exactly this scenario, ninety days instead of thirty is a full month of deficit removed from the calculation.

Arrange credit before you need it. A facility agreed while the business looks strong is available on much better terms than one sought during the quarter you lost a quarter of your revenue. Lenders price on what the business looks like when you ask, and that is a reason to ask early rather than a reason to borrow.

The scope creep problem sits underneath all of this: the client whose fee has stayed flat while the work quietly grew is both less profitable and harder to replace than the number suggests, and how a profitable retainer quietly becomes an unprofitable one is the diagnostic for that. A book full of those is more concentrated in effort than it looks in revenue.

Read it quarterly, in fifteen minutes

Recalculate every quarter, because the inputs move without announcing themselves, and the most common way a business becomes over-concentrated is by growing.

The pattern is ordinary. Your best client is delighted and expands their scope. You are pleased. Revenue is up, the relationship is strong, and the concentration ratio has moved from 22% to 41% over four quarters without a single decision being taken to allow it. Nothing on the P&L flags this; it looks like an excellent year. Concentration risk is one of very few business risks that grows precisely when things are going well, which is why it needs a scheduled check rather than a triggered one.

Four numbers, one page, every quarter:

Largest client as a share of monthly revenue, and separately, as a share of margin.
Top three combined, the same two ways.
Realistic fixed floor, re-honest about who you would actually not let go.
Current replacement time, based on the last two deals you actually closed rather than on how long you would like it to take.

Then re-run the formula. It takes about fifteen minutes, and the value is not the answer in any single quarter but the direction across four. A concentration ratio moving steadily up while the cash balance stays flat is a business becoming quietly more fragile during what feels like its best year, and that pattern is legible a long way in advance, but only to someone who is looking at it. Which is also why the weekly cash view is one of the five things a founder should never fully delegate: the monthly P&L is where this hides, and it hides well. If reading that statement is not yet second nature, start with how to read a profit and loss statement and come back to this.

The one-sentence version: a services business does not need months of runway, it needs enough cash to replace its largest client, and if you have never worked out which of those two numbers is bigger, it is the second one.

How Much Can You Really Make With Affiliate Marketing? An Honest Answer

Prabhash Jha — Sun, 06 Sep 2026 12:52:12 +0000

"How much can I make with affiliate marketing?" is the most-asked and worst-answered question in the space, usually with screenshots and hype. Here's an honest answer, without the fantasy figures.

Most people ask this like effort is the variable. Work harder, publish more, earn more. That framing is why so many put in two years and bank almost nothing.

The variable that decides your ceiling isn't effort. It's how much money one visitor is worth to you. And that number is locked in the moment you pick your niche and your offer, before you write a single word. Two people can publish the same volume, at the same quality, pulling the same traffic. One earns forty times the other. Not because one tried harder. Because one picked an offer paying ₹75 a sale, and the other picked one paying ₹8,400 over the life of a customer.

So the honest answer has two halves.

Part one: most people who try affiliate marketing make very little. A small minority make a lot. It isn't a salary. It's a skill that pays in proportion to the trust and traffic you build. Early on you'll likely earn close to nothing. That isn't failure. That's the normal shape of the curve.

Part two, the part nobody publishes. You can work out your own realistic number in about ten minutes with arithmetic. No screenshots required. No gurus. Just numbers you can pull from a merchant page and a spreadsheet. This post shows you how.

The rest of what you're about to read is deliberately unglamorous. If you were hoping for a story where somebody quit their job and hit ₹5 lakh a month in ninety days, look elsewhere. That story exists. It's just not most people's story, and pretending otherwise is how the whole industry keeps producing beginners who feel stupid for earning ₹400 in month six.

Why there is no useful "average affiliate income"

Every "average affiliate earns ₹X" figure you've seen is built on a survey of people who volunteered to answer, on a platform where affiliate marketers gather. The people earning nothing already quit. They aren't in the sample. The people earning a lot have an incentive to be visible. So you're reading the middle of a filtered slice, not the middle of the population.

Worse, the underlying distribution isn't one where an average means anything. Affiliate earnings follow the same winner-take-most shape as the traffic sources feeding them. In organic search, the top result for a commercial query takes a big multiple of what position five takes. Position eleven takes close to zero. The drop isn't linear. It's a cliff. Affiliate revenue inherits that shape and then multiplies it, because the top-ranked page also gets the best merchant terms, the best conversion rate from a warmer audience, and the leverage to negotiate a bumped commission.

When a distribution is shaped like that, the mean gets dragged upwards by a handful of outliers and describes nobody. The median sits closer to the truth. And the median is low. Quote either one at a beginner and you mislead them.

Why affiliate income reports are the worst data you can learn from

Income reports feel like evidence. Honestly, they're advertising. Five structural reasons they can't be used as a forecast:

Selection. Only winners publish. Nobody writes "month 19: ₹0, here's what I learned." You're seeing the survivors of a process that removes most participants. Then you infer the process works.

The report is the product. The most common way to monetise a big affiliate income report is to sell a course, a community or a tool about affiliate marketing. The screenshot is the sales asset. That doesn't make it fake. It makes it selected, cropped and timed.

Gross, not net. Almost every published figure is gross tracked commission. Before reversals. Before payout thresholds. Before FX and bank charges. Before TDS. Before GST. Before tools and content costs. Before income tax. The banked number can be materially lower. I cover the leakage below.

Unrepeatable conditions. A site that ranked in 2016 did so against a search index, a competitive set and a set of merchant payouts that no longer exist. A media buyer who was profitable at a certain CPM was profitable at that CPM. Copying tactics without the conditions is copying the visible half of the machine.

Volatility is hidden. Monthly screenshots hide the algorithm update that halved traffic. They hide the merchant that cut the commission rate with two weeks' notice. They hide the network that clawed back three months of sales as fraud. Affiliate income isn't annuity income. It's revenue from a channel you don't own, on terms you can't control.

What actually determines how much you earn

Strip the topic to its mechanics. Five levers. Everything else is a tactic that moves one of them.

Your audience's trust. People buy on your recommendation only if they believe you. Trust is the real asset. It's why two sites with identical traffic can convert at rates several times apart.
Traffic and relevance. More of the right people seeing the right offer beats a huge, uninterested crowd. Intent matters more than volume. The gap isn't small.
What you promote. Recurring and higher-value products pay far more than cheap one-off items. This lever has the widest range of the five. It's also the one beginners set carelessly.
Retention on the merchant's side. With recurring commissions, the merchant's churn rate is your income. You don't control it. Most affiliates never check it.
Consistency over time. Affiliate income compounds slowly, then meaningfully. The people who quit early never see it.

Collapse those into one number and you get revenue per 1,000 visitors. Your affiliate RPM. It's the only figure that lets you answer "how much can I make" honestly, because it converts a traffic goal into a rupee goal and back again.

RPM = (visitors who click your affiliate link) × (share of those clicks that become a sale) × (commission per sale), expressed per thousand visitors.

How much traffic do you need to make ₹50,000 a month?

Below is the same question run through five different offer types. The assumptions are illustrative inputs, not benchmarks. The point is the spread between rows, not any single row. In every case I've assumed 8% of visitors click through to the offer, which is a reasonable figure for content written with buying intent and a poor one for general blog traffic.

Offer type	Commission per sale	Click-to-sale rate	Revenue per 1,000 visitors	Visitors/month for ₹50,000
Low-ticket physical (₹1,500 item at 5%)	₹75	4.0%	₹240	~2,08,000
Mid-ticket physical (₹8,000 item at 6%)	₹480	3.0%	₹1,152	~43,000
Digital course (₹6,000 at 40%)	₹2,400	1.5%	₹2,880	~17,000
SaaS one-off bounty	₹3,000	2.0%	₹4,800	~10,400
SaaS recurring (₹600/mo, ~14-month life)	₹8,400 lifetime	1.5%	₹10,080 lifetime	~5,000 (at steady state)

Check the first row so you can trust the rest. 2,08,000 visitors → 16,640 affiliate clicks at 8% → 666 sales at 4% → 666 × ₹75 = ₹49,920, which is ₹50,000 once you stop rounding the visitor count. The last row: 5,000 visitors → 400 clicks → 6 sales a month. At a 14-month average subscription life, the customer base stabilises at 84 paying accounts. 84 × ₹600 = ₹50,400 a month.

Read the two ends of that table again. Same income. Same conversion competence. A 40× difference in the traffic required. If you're grinding for 2,00,000 monthly visitors to a low-commission niche, no amount of extra effort closes that gap. The gap was created by a decision you made in week one.

This is also why "just pick a niche you're passionate about" is incomplete advice. Passion sustains you through the quiet months. It doesn't change your RPM. Pick something you can hold an opinion about and where the money per visitor isn't structurally tiny.

Why recurring commissions beat one-off payouts, with the maths

Take two offers you could promote with the same content and the same traffic, assuming you can steadily generate 24 new referrals a month.

Offer A, one-off: ₹2,000 per sale, paid once. 24 sales × ₹2,000 = ₹48,000 a month, flat, from month one.

Offer B, recurring: ₹600 a month for as long as the customer stays, average life 14 months. Each referral is worth ₹8,400 in total. But the cash arrives in slices.

With a 14-month average life, roughly 7.1% of your customer base cancels each month. Your base stops growing when new additions equal cancellations. 24 = base × 7.1%, so the base settles at 336 customers. They pay you 336 × ₹600 = ₹2,01,600 a month. Four times Offer A. But look at how long it takes to get there.

Month	Share of steady state reached	Monthly income (Offer B)	Monthly income (Offer A)
3	20%	~₹40,200	₹48,000
6	36%	~₹72,400	₹48,000
12	59%	~₹1,18,700	₹48,000
24	83%	~₹1,67,500	₹48,000
36	93%	~₹1,87,600	₹48,000

Offer B's monthly cheque overtakes Offer A around month four. Cumulative earnings cross over around month seven. Over 36 months, Offer B pays roughly ₹48 lakh against Offer A's ₹17 lakh. About 2.8×. On identical work.

Three honest caveats worth naming, because the pretty compounding curve above hides all of them.

First, this is a steady-state model, not a forecast. It assumes you can hold 24 referrals a month from month one, which is exactly the hard part. In practice the first months usually deliver a fraction of that, and the base you're multiplying against grows accordingly.

Second, the whole advantage sits inside the merchant's retention. Which you don't control. If their real average life is 5 months rather than 14, the steady-state base falls to 120 customers and ₹72,000 a month, and Offer A becomes competitive. Ask any recurring programme for their average subscription life or monthly churn before you build content around them. If they won't tell you, assume it's bad.

Third, many "recurring" programmes are recurring for 12 months only, then stop. Read the terms.

And a fourth thing almost nobody checks. Whether the tracking that assigns you those recurring payments actually keeps working. Recurring commissions fail silently more often than one-off ones, because a broken attribution on a subscription costs you fourteen payments, not one. That's worth a separate read: affiliate tracking breaks quietly, and you want to catch it before payout day.

How long does it take to make money with affiliate marketing?

If your traffic is organic, the honest timeline is set by how long search engines take to trust a new site. Not by how fast you publish. Expect months of near-zero while you build content and trust. That's normal. Treating it as failure is the single most common reason people stop three months before it would have started working.

Phase	Roughly when	What is actually happening	What income looks like
Build	Months 0–4	Pages indexed, ranking for almost nothing commercial	Effectively zero
First signal	Months 4–9	A handful of long-tail pages rank; first sales arrive	Sporadic, small, unpredictable
Compounding	Months 9–18	Winners identifiable; you double down and cut losers	Irregular but trending up
Business	Months 18–36	Topical authority; merchants take your calls	Meaningful, still volatile

Paid traffic compresses that timeline and swaps it for a different risk. You can buy visitors today, but you're now buying media against a fixed commission, with no control over the merchant's landing page, checkout, cookie window or approval rate. Your margin is the gap between your CPA and their payout. They can change their side unilaterally. If you go this route, you need to be fluent in the core performance metrics: CPC, CPM, CTR, CPA and ROAS before you spend a rupee. And you need to decide deliberately whether Google or Meta is the right starting platform.

The one lever that genuinely shortens the timeline without adding media risk is owning the audience directly. Search traffic is rented. A list isn't. Building an email list from zero alongside your content means the second, third and fourth offer you promote don't need to be discovered again from scratch.

Which niches pay the most, and what they cost you

The highest per-visitor payouts in India cluster in financial services, insurance, B2B software and high-ticket education. Not a coincidence. Payouts track the lifetime value of the customer to the merchant, and those customers are worth a great deal.

They're also the hardest places to operate, for two separate reasons.

Competition. High payout attracts capital. You're competing against comparison sites with full-time editorial teams and direct commercial relationships. Not against other beginners.

Regulation, which most guides ignore entirely. In India, insurance isn't an open affiliate category. Soliciting or procuring insurance business for a commission requires registration with IRDAI as an intermediary. Agent, broker, corporate agent or an approved digital arrangement. Putting an insurance affiliate link on a blog and taking commission without that registration isn't a grey area. If insurance offers appear in your network dashboard, check the structure before you promote them.

Investing and broking referrals sit under SEBI's rules. You can explain how a product category works. You must not give personalised investment advice or recommend specific stocks or securities without being a registered adviser or research analyst. And referral arrangements for broking have their own restrictions. Educational content is fine. Recommendations are a different legal object entirely. Nothing on this site, including this post, is investment advice.

Separately, and regardless of category. ASCI's influencer guidelines require you to disclose a material connection prominently wherever you promote for commission. A buried "this page may contain affiliate links" at the footer isn't a disclosure. Put it where the reader sees it before they click. It costs you almost nothing in conversion. And it's the difference between a durable asset and one that gets taken down.

If you're still choosing, work through a proper checklist for picking an affiliate programme rather than signing up for whatever pays the loudest.

Gross commission is not take-home

The number in your network dashboard is tracked commission. Several things sit between it and your bank account. Beginners routinely plan around the wrong one.

Reversals and clawbacks. You're paid on approved sales, not tracked ones. Returns, cancelled subscriptions, duplicate orders and fraud-flagged conversions come back out. Approval rates vary by category. They can be brutal in e-commerce.
Payout thresholds and terms. Most networks pay net-30 to net-60 after month end and hold your balance until it crosses a minimum. Your first real payment can land four months after your first sale.
FX and bank charges. Payouts from foreign networks arrive in USD and lose something to the conversion spread and inward remittance charges.
TDS. Commission paid by an Indian payer is generally subject to tax deducted at source under section 194H. It's a credit against your tax, not a loss. But it reduces the cash that arrives.
GST. Affiliate promotion is a supply of services. Registration thresholds apply. Commission earned from a foreign network may qualify as export of services with its own treatment. Get this checked by a CA before your income is large enough to matter. It's far cheaper to set up correctly than to unwind.
Costs. Hosting, email tool, any paid content or design, and the tracking you need to know what's working.

Plan on a meaningful gap between tracked and banked. Then income tax on top. If you're running this as a business rather than a hobby, understand why cash flow and profit are not the same thing. Net-60 payout terms are a cash flow problem long before they're a profit problem.

What separates the people who earn from the people who don't

Not talent. Not hours. Four things, in order of impact.

They chose an offer with survivable economics, so their traffic goal was 20,000 a month rather than 2,00,000. They wrote for people who were already deciding what to buy, not for people idly curious about a topic. They tracked which pages produced approved sales and killed the rest without sentiment. And they stayed long enough for compounding to happen. Which, given the timeline above, means committing eighteen months before judging the result.

The reverse pattern is equally consistent. People fail because they quit during the long, quiet early phase. Or they chase too many things at once. Or they promote products they don't believe in. Focus, patience and honesty are what separate the ones who earn.

A realistic path, not a jackpot

Do the arithmetic first. Pick a target monthly income, find offers in your area with a real commission structure, and calculate the traffic you'd need. If the answer is 2,00,000 visitors a month, change the offer, not the target.
Expect months of near-zero while you build content and trust. Budget for it in time and money so the quiet phase doesn't force you out.
Pick one niche and a few good programmes instead of promoting everything. Depth in one topic is what earns rankings and trust. Breadth earns neither.
Track what converts right down to approved sales by page, not clicks. A simple sheet that tells you whether your marketing makes money is enough to start.
Reinvest your attention into what's working, and be patient. Most of your income will come from a small number of pages. Your job is to find them and feed them.

FAQs

How much can a beginner make in affiliate marketing in the first year?

Plan for close to nothing in months one to four. Small and irregular amounts through months four to nine. Something meaningful only if your offer economics are strong and your traffic has commercial intent. Anyone promising a specific first-year figure is selling something. The honest answer depends entirely on which offer you chose.

Is affiliate marketing still profitable in 2026?

Yes, but the easy version is gone. Thin comparison pages and coupon scraping no longer rank or convert. What still works is genuine depth on a narrow topic, aimed at people at the point of decision, with offers that pay enough per customer to justify the traffic you can realistically build.

Can affiliate marketing replace a full-time income?

For some people, eventually. But it's a minority, and it usually takes years of consistent work, not weeks. Treat early income as proof of concept, not a pay cheque. If it does replace your salary, remember it's revenue from a channel you don't own. Build a second traffic source before you resign.

Why do most people fail at affiliate marketing?

They quit during the long, quiet early phase, chase too many things at once, or promote products they don't believe in. Structurally, most also pick offers where the commission per visitor is so low that the traffic required was never achievable. Focus, patience and honesty are what separate the ones who earn.

How much traffic do I need to make ₹1 lakh a month from affiliate marketing?

Divide ₹1,00,000 by your revenue per 1,000 visitors. On the ₹1,152 RPM from the mid-ticket row above, that's roughly 86,800 visitors a month. On the ₹4,800 RPM of the SaaS bounty row, about 20,800. Work out your own RPM before setting any traffic goal. The goal is meaningless without it.

Are affiliate income reports and screenshots real?

Usually the numbers are real and the conclusion you draw from them is wrong. They show gross tracked commission from a survivor whose conditions you can't reproduce, published by someone who often earns from teaching affiliate marketing. Use them for tactics if you like. Never use them as a forecast for yourself.

Key takeaways

Your earnings ceiling is set by revenue per 1,000 visitors, and that number is decided when you choose your offer, not by how hard you work afterwards.
The same ₹50,000 a month can require 5,000 visitors or 2,08,000 visitors depending on the offer. A 40× difference on identical effort.
Most people earn little. A focused, patient minority earn a lot. The distribution's shape makes any "average income" figure meaningless.
Income reports are survivorship bias sold as evidence. Gross, selected, timed and produced by people who monetise the telling.
Recurring commissions can pay several times more than one-off ones over three years, but only if the merchant's retention is real. Ask for their churn before you commit.
Trust and relevant traffic decide your income. Expect months of near-zero. And know that gross tracked commission isn't what reaches your bank account.

How a Profitable Retainer Quietly Becomes an Unprofitable One

Prabhash Jha — Sat, 05 Sep 2026 12:41:36 +0000

The retainer was profitable in month one. By month six it was quietly "under review", which is the polite phrase agencies use for the moment the finance meeting stopped being fun. Nothing had changed on paper. Same fee. Same deliverables. And yet the account was quietly losing money every week, and the founder running it didn't notice until the quarterly P&L was already printed.

This post is about how that happens. More usefully, it's about the two costs that cause it. Costs that never appear on any scope document. The monthly measurement that catches them before the P&L does. And the re-scope conversation that saves the account instead of ending it.

The advice that misses

Type "how to prevent a retainer becoming unprofitable" into a search engine and page one is roughly the same everywhere. Track hours per client. Flag the account when time-spent hits 75% of the budget. Have a monthly reconciliation call. Charge for out-of-scope work.

All fine advice. It's aimed at scope creep in the visible sense, the sort where a client asks for an extra landing page and you build one. For that failure mode it works.

What it doesn't catch is the failure mode most retainers actually die from. Which isn't new tasks the scope document didn't cover. It's old tasks the scope document doesn't describe accurately, done more expensively than the fee assumed, in ways that will never show up on a timesheet because nobody thinks to log them.

The two costs doing the killing are response latency and unbilled thinking time. Both invisible to the industry-standard tools. Both compound weekly. Both fixed by the same re-scope conversation script, the one below, which no vendor is going to publish because the vendor sells the dashboard that doesn't see them.

Cost one: response latency

Response latency is what it costs you when a client's questions get answered too quickly.

Every retainer starts with an unspoken assumption about how fast the client can expect a reply. If you never make it explicit, the client sets the default themselves. And the default they set is: whenever they email, whenever they message, whenever they call.

In month one, when there are three other clients and the account is fresh, replying inside the hour isn't expensive. By month six, when there are eleven other clients and this one has grown a habit of pinging you at 4pm on a Wednesday about a screenshot they took at 3:55pm, the same hour-inside response is expensive in a very specific way.

Here's how. It fragments the working day of your most-billable people. A senior producer who could have done four hours of concentrated work in an afternoon can instead do six twenty-minute stretches broken up by two-minute Slack replies. The difference between those two afternoons is the difference between an account that runs on margin and one that runs on goodwill.

The client hasn't asked for anything the scope document doesn't describe. They've asked for a screenshot in a Slack message. The cost of the reply isn't the two minutes it took to send. It's the twenty minutes on either side of it, which is the well-documented cost of interrupted work.

Almost no agency measures this. Not because they don't know it's happening. The producers know exactly when it's happening. The tooling to measure it doesn't exist and, honestly, the client would find the measurement embarrassing. So the cost stays invisible, and the account owner reports on hours logged and deliverables shipped while the margin drips out of a hole nobody has named.

The other thing worth saying about response latency is that the client is not the villain of the story. The client has been trained by the agency's own habits. A reply inside the hour in week two is a promise the client has no reason to disbelieve, and by week eight it is the working assumption on both sides. What looks like a client behaviour is actually a reinforcement schedule the agency set up itself. Which is useful, because it means the fix is on the agency side. You can rewrite the schedule. You can't rewrite the client.

Cost two: unbilled thinking time

Unbilled thinking time is what it costs you to do the work that happens before the work you're billing for.

Every reasonable retainer includes deliverables. A monthly report. A set of creatives. A campaign build. A strategy note. The scope document is written around those. What the scope document doesn't say is that most of them take about half as long to produce as they take to think through, and the thinking-through is what actually eats the account manager's brain.

Take the "monthly performance review" that appears on almost every marketing retainer. The document itself, once you know what you want to say, takes maybe two hours to write. The thinking-through takes six. Reading the last four weeks of dashboards. Running the diagnostics. Working out what actually changed versus what's noise. Framing it in a way the client will trust.

Nobody logs those six hours anywhere. On the timesheet, the account is booked for "monthly report, 2 hours". On the payroll, the account manager has spent Tuesday, Wednesday morning, and part of Thursday morning thinking about this client. The retainer was priced for the timesheet. The account is being delivered by the payroll.

Multiply that gap across ten clients and it isn't a gap anymore. It's the entire margin.

There's a second-order version of this that's worth naming. The account manager who has been spending six unlogged hours a month thinking about the client also becomes the person the client trusts. So the next month, when the client asks for a piece of ad-hoc thinking that wasn't in the scope, they ask that account manager, not the wider team. Which means the six hours becomes eight, then nine. The account gets better outcomes and the retainer gets worse economics. Both things are true at the same time, which is why the trend is easy to miss from the top.

The compound

The two costs aren't additive. They're multiplicative.

A client who has trained you to reply within the hour is also a client whose account manager can't get any thinking time done during the working day. So the thinking-time cost, the six unlogged hours per month, moves to evenings and weekends. Where it's done tired, badly, by people who will eventually resign. The retainer looks fine on paper for another quarter. The turnover in the team is where the cost actually lands.

Meanwhile the client's own perception of value has been quietly recalibrated by the response latency. A retainer whose deliverables ship monthly but whose replies arrive within the hour trains the client to see the retainer as an on-call resource. They stop remembering the deliverables. They start remembering the availability. When it's time to renew, the negotiation isn't about the deliverables you've shipped. It's about the availability they've come to expect. Next month's fee gets priced against the deliverables. Next month's cost gets delivered against the availability.

The monthly measurement that actually catches this

You don't need a timesheet. The people who fill in timesheets have been trained to fill them in against the scope document, so timesheets tell you exactly what the scope document already told you.

What you need is a two-column note per client. Twenty minutes at month-end. Done by whoever owns the account. It looks like this.

Column one: shipped this month. The list of things that came off the account this month. Deliverables, reports, campaigns, strategy notes. Anything the client received. Plain language, one line each.

Column two: everything else this month. Everything the account team did that didn't appear in column one. Every meeting. Every Slack conversation over ten messages long. Every email chain over five replies. Every ad-hoc analysis. Every time somebody redid a piece of work because the first version got sent back. Every 4pm-on-Wednesday screenshot request. Plain language, one line each.

You're not measuring hours. You're counting entries. If column two has more entries than column one, the retainer is running on unbilled thinking time and response latency, and the numbers on the P&L are lagging what's already happened.

Why this works when timesheets don't: it doesn't ask anyone to remember how long something took. It only asks whether it happened. That's a question a human being can honestly answer at month-end. How long something took is a question they can't.

The three warning signs before the P&L catches up

Do the two-column note for three months and the pattern will announce itself before the finance meeting does. The warnings, in the order they show up:

Sign one: the account manager stops proposing anything. In month one, the account team was full of "we should try X". By month six, when the same team is asked what to try next, they say "we should keep doing what we're doing". This isn't agreement with the strategy. It's the operational reality that they don't have any thinking-time left. A retainer whose team has stopped proposing new work is a retainer delivering only from muscle memory, which is the most expensive way to deliver anything.

Sign two: the client's Slack messages start with "quick one". The word "quick" in a client message is almost never accurate. What "quick one" actually means is: this is a small enough thing that I don't feel bad asking, but a specific enough thing that I need a real answer. The client isn't being manipulative. They've simply learned that "quick one" gets a response inside the hour. Once the pattern establishes itself, you've become a shared inbox for their operations without noticing.

Sign three: deliverables start slipping and nobody knows why. The monthly report is a day late, then two, then three. The team isn't idle. The team is exhausted. Nobody is doing less work. They're doing less work that's on the scope document, because everything not on the scope document has to happen first, in real time, at 4pm on a Wednesday.

When you see all three signs on the same account, the P&L reconciliation for that account is going to be ugly. You have between four and eight weeks. Have the conversation now.

The re-scope conversation

The reason this account is losing money isn't that the fee is wrong. It's that the delivery model is wrong. The conversation you need to have is the one that fixes the delivery model. The reason most agencies never have it is they think the conversation is a request for more money. It isn't. It's a request for the account to survive.

Here's the shape of the conversation, in three moves. Each move is one sentence long, in normal working language, said by the account owner and heard by whoever holds the budget on the client side. This isn't a script to be recited. It's the sequence and the specificity.

Move one: name the pattern, not the client's behaviour. The account owner opens with a sentence like: "This retainer was built around monthly deliverables, but the work has moved toward real-time back-and-forth, and I want to reset how we run it before the next quarter, so the delivery model matches how we're actually using each other." Notice what this doesn't say. It doesn't blame the client for asking too much. It doesn't say the retainer is unprofitable. It names the pattern and asks to re-scope. Both neutral facts.

Move two: propose the two structural changes, in a form the client can accept without losing anything. The two changes are almost always the same. First: response windows. Slack replies within one working day rather than one hour, urgent things flagged as urgent, and one weekly office-hours block for real-time questions. Second: strategy time. A named half-day per month for the account team to think through what's next, protected from the ongoing work. Neither change costs the client anything in absolute terms. Both save the account.

Move three: offer the trade. Where a fee increase is warranted, this is where you name it, and you name it in exchange for something specific. A new deliverable, an expanded reporting cadence, a stated ambition. Where a fee increase isn't warranted, this is where you propose the delivery-model reset without a fee change and offer to review the fee in the following quarter if the model has settled. The important thing is that the trade is specific. "We need to raise the fee" is a fight. "We are moving to a model where we can protect strategy time; here is what that adds" is a conversation.

Have the conversation with whoever holds the budget. In person or on video. Never by email. Fifteen minutes on a calendar invite booked two days in advance is enough. If you can't get the meeting, you've already found out something about the account that's worth acting on separately.

What to do if the conversation fails

Some accounts won't accept the re-scope. When that happens you have three options, and the one you pick is a business decision, not an operational one.

Option one: raise the fee to match the delivery model. If the client wants to keep responding in the hour and doesn't want a named half-day for strategy, the price for that is a fee that assumes the higher cost of delivery. Model it honestly. If the current fee is X and the account is being delivered at cost 1.3X, the fee that makes it profitable at the current delivery model is closer to 1.6X than 1.3X, because you also need margin. Present the number with the reasoning. Be prepared for the client to leave. That's a fair outcome. The mechanics of the announcement itself, which accounts to raise on first, which to raise on last, and the meeting move that decides whether the increase actually holds, are in how to raise your prices without losing the accounts you actually want to keep; the order matters more than the wording.

Option two: reduce the deliverables to match the fee. If the fee can't go up and the delivery model can't change, the honest response is to reduce what you ship each month so the account's actual cost matches the actual revenue. Drop the monthly report to a quarterly one. Drop the strategy note. Whatever you drop, name it, and get the client's written acknowledgement that this is now the scope. Unpopular. Honest.

Option three: end the retainer at renewal. If neither of the above works, the account is unprofitable at any delivery model the client is willing to accept, and the right move is to let it end at renewal, on good terms, with a handover that protects the client. Don't fire the client mid-quarter. Time it to a natural boundary and offer to introduce them to two competitors who could serve them well. This costs you nothing that was not already lost, and it saves the reputation on both sides.

The one option that's not on the list is "continue and hope". The account isn't going to fix itself. Every month you don't act, the two-column note gets more lopsided and the account manager gets closer to a resignation letter.

Why this is not a dashboard problem

The temptation, having read the above, is to buy a piece of software that measures response latency and thinking time. There are many such pieces of software. They won't solve the problem, for two reasons.

The first is that measurement without a conversation is just data. If the account owner sees a red number on a dashboard once a month and doesn't have a script for what to do about it, the red number becomes background. The sort of thing you learn to notice without acting on. The two-column note works because it's done by the person who has authority to act, in the same twenty minutes it's done in. There's no lag between measurement and action.

The second reason is that the thing that fixes the account isn't the data. It's the willingness to have the conversation. Every account owner who has ever run a retainer knows, without a dashboard, which of their accounts are running on goodwill. The measurement isn't what surfaces it. The measurement is what gives them cover to raise it. If you're the founder, your job isn't to buy them the measurement. It's to give them cover, and to have the conversation yourself on the accounts where they can't.

Where this fits with the rest of the work

The retainer economics conversation sits next to a small number of related conversations, and understanding all of them together is more useful than understanding any of them alone.

If your problem is that you can see the account is unprofitable but you can't see the cash consequence yet, the difference between what you're measuring and what your bank balance is measuring is covered in Cash flow vs profit: the difference that sinks most small businesses. A retainer can be loss-making for six months before the cash-flow statement notices, and the reverse is also true. Don't conflate the two.

If your problem is that this isn't one drifting account but a whole book of them, and you can't tell which client the drift is actually in, the cross-sectional version of this measurement is in which client is actually profitable when everyone works on everything. How to allocate a shared team's cost across accounts without a timesheet culture.

If your problem is that you haven't built a retainer at all yet and you're pricing services one project at a time, the framework for setting a rate that protects against exactly the failure mode above is in How to price your services as a freelancer or consultant. Pricing badly at the start makes the re-scope conversation harder for years.

If your problem is that a client has already stopped paying, not is-not-quite-profitable but is-not-paying-at-all, the response is a different sequence and a different set of levers, in The client has stopped paying: here is the sequence, in order.

FAQ

How long does it usually take for a profitable retainer to turn unprofitable?

In most cases it happens between month three and month six, which is roughly when the delivery habits harden into a pattern the fee was never priced for. You won't see it on the P&L for another quarter after that. Which is why the two-column note at month-end matters more than the finance meeting.

Should I raise the fee or reduce the deliverables when an account slips?

Raise the fee when the client values what they're actually getting. The availability, the real-time replies, the informal strategy. Reduce the deliverables when they only value what's written on the scope. If you can't tell which one it is, the re-scope conversation is what tells you, because their reaction to the two structural changes is the answer.

What if the client refuses the response-window change?

That refusal is the answer to a question you were already asking. Which is whether this account can ever be profitable at the current fee. Move to option one from the post, raise the fee to match the delivery model, and if they refuse that too, time the exit to renewal.

How do I bring this up without sounding like I am asking for more money?

Open with the pattern, not the fee, and propose the two structural changes before you propose any commercial change. If the account genuinely needs a fee increase, it becomes the third move in the conversation, not the first, and by then you've already reframed the meeting from a negotiation into a working session.

Do I have this conversation with every client at once, or one at a time?

One at a time, starting with the account whose two-column note is most lopsided, because that's the one where the argument is easiest to make and the outcome is most instructive. Once you've run the conversation twice, you'll have a version of it that sounds like you rather than like a script, which is what the third client will need.

What if the account manager is the founder, is the conversation different?

The mechanics are identical, but the discipline is harder, because the founder has more tolerance for absorbing the invisible costs personally and calls that tolerance loyalty. The two-column note is more important, not less, when the person delivering the account is the person who will not resign over it.

Is a shared Slack channel with the client always a bad idea?

No, but a shared channel without a stated response window is, because the channel itself sets the expectation of an inside-the-hour reply. Keep the channel and add the window. One working day for normal messages, an "urgent" flag for genuine urgency, and a weekly office-hours block for real-time back-and-forth.

The one-sentence version

A retainer becomes unprofitable when the delivery model outgrows the scope document. The delivery model outgrows the scope document by two things, response latency and unbilled thinking time, and both are invisible to the tools that measure hours. The re-scope conversation is what turns the invisible into a decision, and the decision is what saves the account.

Do the two-column note this month. Have the conversation before you present the P&L. It's a shorter meeting, and a much cheaper one, than the one where you tell the team you're letting an account go.

Google Ads vs Meta Ads: Which Should a Beginner Start With?

Prabhash Jha — Sat, 05 Sep 2026 12:41:35 +0000

Almost every answer to this question is a feature comparison. Google has search. Meta has targeting. Both have video. Here's a table. Done.

That answer misses the point completely. Google Ads and Meta Ads aren't two versions of the same product. They do opposite jobs. Which one you start with comes down to a single question about your business.

Does anyone already know they need what you sell?

If yes, start with Google. If no, start with Meta. Everything below is the reasoning, the exceptions, and how to run whichever you pick without burning the first month.

The difference that matters: intent

Google Ads is demand capture. Someone types "emergency plumber near me" or "best project management software for small teams". They have a need. They're actively looking. Your ad is an answer to a question they already asked. You aren't creating the demand. You're competing to serve it.

Meta Ads is demand creation. Nobody opens Instagram to buy anything. Your ad interrupts something they were enjoying, so it has to earn attention before it can sell. You're either creating awareness of a need, or reminding someone of one they'd forgotten.

One thing worth knowing before you compare their reported numbers. Demand capture is far easier to measure than demand creation. Budgets drift toward whatever is easiest to count. That drift is one of the bills digital advertising quietly charges, and it makes the capture channel look better than it is on any dashboard that treats the two as equivalent.

Almost every practical difference falls out of that one distinction:

	Google Ads	Meta Ads
The user is	Searching	Scrolling
You are	Answering	Interrupting
Intent	High	Low to none
Cost per click	Higher	Lower
Conversion rate	Higher	Lower
Creative matters	Somewhat	Enormously
Time to first sale	Faster	Slower
Scales to new audiences	Limited by search volume	Very well
Good for	Known problems	Unknown or visual products

Look at the cost lines. Google clicks cost more and convert better. Meta clicks cost less and convert worse. Neither is cheaper. They just distribute the cost differently. Comparing cost per click between them tells you nothing useful.

Start with Google if…

People search for what you sell. Test it yourself. Open Google. Type what a customer would type. If there are ads already running against that query, someone is paying for those clicks, which means the demand is real and monetisable. Honestly, that's the single most useful thirty seconds of research available to a beginner.

Your product solves an urgent or obvious problem. Plumbing. Legal help. Accounting software. Replacement parts. Medical services. People go looking.

You have a small budget and need results soon. Google's learning period is shorter. Traffic is qualified from day one. You'll get fewer clicks for your money, but more of them will matter.

Your product is hard to explain visually. B2B software, professional services, anything where the buyer is researching rather than browsing.

The catch: you're limited by how many people search. If only 200 people a month search for what you do, Google can't give you more than that. Doesn't matter what you spend. This is the ceiling that surprises people. You can max out search demand and simply have nowhere else to put budget.

Start with Meta if…

Nobody is searching, because they don't know it exists. New product categories. Novel solutions. Impulse purchases. If your keyword research finds no volume, that's not a dead end. It means the demand has to be created.

Your product is visual. Fashion, food, interiors, fitness, anything where seeing it is most of the sell.

You know your audience demographically or behaviourally, even if they aren't searching. Meta's targeting is genuinely strong here, though less precise than it was before the privacy changes of recent years.

You want to build an audience, not just capture sales. Meta is far better at reaching people who've never heard of you, and at re-reaching them cheaply afterwards.

The catch: creative carries the entire campaign. A bad ad on Meta doesn't underperform. It fails completely. You'll need several versions. You'll need to keep making new ones, because performance decays as an audience sees the same thing repeatedly. If you can't produce creative regularly, Meta will be frustrating.

Realistic budgets

The most common way beginners waste money is spreading too little across too much.

Minimum to learn anything: roughly ₹15,000–25,000 (about $200–300) over a month, on ONE platform. Below that, you won't gather enough data to know whether it worked. You'll draw conclusions from noise.

Do not split a small budget across both. Two half-funded campaigns learn slowly on both platforms and teach you nothing about either. One platform for three months beats two for six weeks.

Expect the first month to be tuition. You're buying data. Which keywords convert. Which creative holds attention. Which audience responds. Judging month one on profit is how people quit two weeks before the thing starts working.

Set a daily budget you could lose entirely without it hurting. Because you might. And the alternative is making panicked changes that destroy the learning.

What people get wrong on each

Google Ads

Bidding on broad keywords. "Marketing" will burn your entire budget on people who wanted a marketing degree. Start with specific, long-tail phrases. Lower volume. Far higher intent. Much cheaper.

Ignoring negative keywords. This is the single biggest waste in most beginner accounts. Add "free", "jobs", "salary", "course", "DIY" and anything else that signals someone who'll never buy. Review the actual search terms report weekly and keep adding.

Sending all traffic to the homepage. The ad promised something specific. The homepage delivers everything. Match the landing page to the ad, or you're paying for clicks that bounce.

Leaving automatic settings on without understanding them. Google's defaults are optimised for Google's revenue, not yours. Check what campaign type you're actually in. Check where your ads are actually showing.

Meta Ads

One ad and hoping. You need three to five distinctly different creatives. Not five colour variations of one idea. Different hooks. Different angles. Different formats.

Judging too early. Meta needs volume and time to optimise. Killing an ad set after two days means paying for the learning phase and discarding what it learned.

Targeting too narrowly. The instinct is to define a tiny perfect audience. In practice, Meta's algorithm now performs better with broader targeting and strong creative. Let the creative do the qualifying.

No landing page discipline. Traffic from an interruption is colder than traffic from a search. The page has to restate the offer, not assume they remember it.

What each platform actually contains

Both are umbrellas over several very different products. Beginners routinely end up in the wrong one without realising.

Google Ads is not just search.

Search. Text ads against a query. This is what people mean when they say Google Ads. It's where beginners should stay.
Performance Max. An automated campaign spending across search, YouTube, Gmail, Maps and the Display Network at once. Google pushes it hard. It works with enough conversion data to learn from, and burns budget opaquely without it. Not a first campaign.
Display. Banners across millions of sites. Very cheap clicks. Very low intent. A large share of the inventory is worth nothing. Useful for retargeting, poor for cold traffic.
YouTube. Video, closer to Meta in behaviour than to search.
Shopping. Product listings for e-commerce. If you sell physical products, this often outperforms text search.

The trap: accepting Google's suggestion to enable Display alongside Search. Your budget quietly drains into cheap low-intent clicks. The blended numbers look fine while search performance is being subsidised by nothing.

Meta Ads spans placements too. Facebook feed, Instagram feed, Stories, Reels, Marketplace. You generally don't pick these manually. The algorithm distributes. What matters is that creative has to work in the format where it lands. A horizontal image built for a feed performs badly as a full-screen vertical Reel. A lot of "Meta didn't work for us" is really "we uploaded one landscape image".

For a first campaign: Google Search only, or Meta with vertical-first creative. Everything else can wait until you know your numbers.

Measure both the same way

Whichever you choose, decide these before spending anything:

What is a conversion? Purchase. Qualified enquiry. Booking. Not a click, not a view.
What is it worth in margin, including repeat purchases?
What can you afford to pay for one? That is your target cost per acquisition. It's the number every decision is judged against.

Without those, you'll optimise toward cheap clicks. Which is the most common way to spend money confidently and get nothing. Marketing metrics explained covers the vocabulary you'll be billed in.

Two measurement notes that trip people up.

Attribution differs between the platforms. Meta claims conversions it influenced within its own attribution window. Google claims ones it influenced within its. Add both numbers together and you'll "achieve" more sales than you actually made. Trust your own order count, not the platform dashboards.

Neither platform is a neutral reporter. Both are incentivised to show you their contribution generously. That isn't dishonesty, it's how attribution works. But it means the comparison has to happen in your accounts, not theirs.

A sensible first 90 days

Weeks 1–2: decide and prepare. Run the search test. Pick one platform. Write down what a conversion is worth. Build one landing page that matches what the ad will promise.

Weeks 3–6: launch small. One campaign. A handful of keywords or three to five creatives. Change one variable at a time. Resist the urge to fix it daily.

Weeks 7–10: cut and double. Kill what clearly isn't working. Move the budget into what is. This is where most of the improvement actually happens, and where most people instead start a new experiment.

Weeks 11–12: decide honestly. Is the cost per conversion below what you can afford? If yes, scale slowly. If no, the problem is more likely the offer or the landing page than the platform. Switching platforms won't fix either.

The honest answer

Most beginners should start with Google Ads, for one unglamorous reason. It fails faster and more legibly. If people are searching and you still can't convert profitably, you learn something true about your offer or your pricing within weeks. Meta can take months to reach the same conclusion, because a poor result is always ambiguous. Was it the creative? The audience? The offer? Not enough time?

But if nobody is searching for what you sell, Google can't help you at all, and starting there is a month wasted proving the obvious.

Run the search test first. It answers the question better than any comparison article. Including this one.

What almost certainly should not be on the shortlist yet is programmatic display. Both platforms here sell you inventory they own, so there's one counterparty and one place the money goes. Programmatic routes your budget through a chain of intermediaries to publishers you've never heard of. Learning to audit that chain is a separate skill from learning to run ads. Learn the ads first.

FAQs

Is Google Ads or Meta Ads better for beginners?

Google Ads, for most people. It captures existing demand, so the traffic is qualified from day one and results appear faster. The exception is when nobody searches for what you sell, in which case Google has no demand to capture and Meta is the only option that works.

Which is cheaper, Google Ads or Meta Ads?

Neither. Google clicks cost more and convert better. Meta clicks cost less and convert worse. Comparing cost per click between them is meaningless. Compare cost per conversion against what a customer is worth to you in margin.

How much budget do I need to start?

Around ₹15,000–25,000 (roughly $200–300) over a month on one platform. Below that you won't collect enough data to distinguish signal from noise. Splitting a small budget across both platforms means learning nothing on either.

Can I run Google Ads and Meta Ads at the same time?

Eventually, and it works well. Meta creates awareness. Google captures the search that follows. But not at the start. One platform run properly for three months teaches more than two run casually for six weeks.

Why are my ads getting clicks but no sales?

Almost always the landing page or the offer, not the ads. Clicks prove the ad worked. If people arrive and leave, check that the page delivers exactly what the ad promised, that the price and offer are competitive, and that the next step is obvious.

How long before ads start working?

Expect the first month to be tuition. You're buying data, not profit. A reliable read takes around 90 days. Launch small, change one variable at a time, then cut what fails and double what works.

Key takeaways

The deciding question is whether anyone is already searching for what you sell.
Google captures demand. Meta creates it. They aren't alternatives, they're different jobs.
Neither is cheaper. Google costs more per click and converts better.
One platform, one budget, three months. Splitting a small budget teaches you nothing.
On Google, negative keywords are the biggest saving available. On Meta, creative is the entire campaign.
Trust your own order count, not the platform dashboards. Both over-claim by design.

GST and TDS Both Hit Your Cash Before the Client Does

Prabhash Jha — Fri, 04 Sep 2026 13:31:36 +0000

If your services business is profitable on paper and short of money in the bank, the cause is usually not bad clients or bad pricing. It's simpler than that. The Indian tax system collects from you on the invoice date. Your clients pay you on their own schedule, which is a different schedule entirely. Nobody sat you down and explained that those are two different calendars, and that the difference between them is where the money you thought you had actually lives.

This post is about that gap, and about the working-capital plan that closes it. It's the version I wish someone had drawn for me on a whiteboard the first time I ran a services business with staff to pay and a receivables ledger that looked healthy right up until the 20th of the month, and then very much did not. The tax portals cover the compliance mechanics well. They'll tell you which form, which date. What they don't cover, because it isn't what they sell, is the operational consequence. How much cash you actually need in the account to run a business that's doing perfectly well. And what to change so that number gets smaller.

Everything below is general information about how the mechanics work. It isn't tax advice for your situation. Rates and thresholds change, sometimes mid-year, and sometimes with retrospective effect. Your CA is the person who signs off on what applies to you. This is the thing to walk in and ask them about, with the numbers from your own P&L in hand rather than a generic query.

Why is my business profitable but out of cash?

Because profit is measured on the invoice date and cash is measured on the payment date. And in a services business in India there are three separate events that pull money out before the client's payment ever arrives.

Most founders understand the first-order version of this. The difference between cash flow and profit is the single most common thing small businesses get wrong. What's less understood is that Indian indirect and direct tax both make the gap structurally worse, in ways specific and predictable enough to plan around.

Here's the sequence, in the order it actually happens:

You raise an invoice. The revenue lands in your P&L now. Your GST liability is created now.
Your GST becomes payable on the 20th of the following month, whether or not the client has paid you.
You pay salaries, rent and vendors on their own fixed dates, none of which are linked to client payments.
The client pays you 30, 45, or 60 days later and pays you less than the invoice, because they've deducted TDS.
The TDS deducted is your money, but you can't touch it until you file your income tax return and the refund is processed, which is the following financial year.

Four of those five events happen before the money arrives. That's the whole problem in one list.

When exactly does GST become payable, on the invoice or on the payment?

On the invoice. In almost every case that matters to a services business.

Under section 13 of the CGST Act, the time of supply for services is the earlier of the invoice date or the date payment is received. Since you almost always invoice first, the invoice date is what triggers the liability. Most services are taxed at 18%. That tax is then paid through GSTR-3B, which for monthly filers is due on the 20th of the following month.

Read that again with a calendar in front of you. The worst case is worse than it sounds. An invoice raised on 1 September creates a liability payable by 20 October. That's 49 days. An invoice raised on 30 September creates a liability payable by 20 October. That's 20 days. Same month, same client, less than three weeks of runway on one and seven on the other, purely because of where in the month the invoice fell.

Now put a 45-day payment term on top. The 30 September invoice isn't due from the client until mid-November. You'll have paid the government the tax on that invoice roughly a month before the client pays you.

You're not paying it out of that client's money. You're paying it out of somebody else's, most likely a client who paid earlier and whose fee was supposed to fund the next month of salaries.

The 18% is not yours to spend. The single most expensive mistake I've watched small services businesses make is treating the GST component of a received payment as revenue, because it landed in the same bank account on the same day as the fee did. It's not revenue. It's a collection you're making on behalf of the government, remitted on a fixed date. A business that spends it is running on a float that has a due date. The day it can't cover the float is the day it discovers it's been insolvent for six months.

If late payment happens anyway, the cost is interest at 18% per annum on the outstanding tax, plus a per-day late fee. That's not a penalty rate designed to be survivable as a funding source. It's more expensive than almost any working capital line you could get instead.

One exception worth knowing if any of your clients are overseas: export of services is zero-rated, so the 18% timing trap above doesn't apply to those invoices — but it's replaced by a different clock, because the zero rating depends on being paid in convertible foreign exchange within a defined period. FX exposure for Indian service businesses covers that clock alongside the currency risk that comes with it.

How much less than the invoice actually arrives?

Meaningfully less. The arithmetic is worth doing once, properly, on a round number.

Take a straightforward professional services invoice. The fee is ₹10,00,000 and GST at 18% is charged separately on the invoice, as it should be.

Line	Amount
Professional fee	₹10,00,000
GST @ 18%	₹1,80,000
Invoice total	₹11,80,000
Less: TDS @ 10% under s.194J	−₹1,00,000
Cash actually received	₹10,80,000
Less: GST remitted to government	−₹1,80,000
Cash you keep	₹9,00,000

The invoice says ₹11,80,000. The amount that survives to fund salaries and rent is ₹9,00,000. That's 76% of the invoice value. And 90% of the fee.

Two mechanics are doing the work there. Both worth naming:

TDS is deducted on the fee, not on the total. Under section 194J, TDS on professional fees is deducted at 10%, with technical services at 2%, and it's deducted on the value excluding GST, provided the GST is shown separately on the invoice. That last clause isn't optional decoration. CBDT Circular 23/2017 is what establishes it. If your invoice format bundles the tax into a single inclusive figure, the client is within their rights to deduct TDS on the whole ₹11,80,000 instead of on ₹10,00,000. On this invoice that's a difference of ₹18,000 of your cash, locked up for a year, entirely because of how the invoice was laid out.

Show GST as its own line. Every time. It's the cheapest working-capital improvement available to a services business and it takes one change to an invoice template.

If your client is a government department or a PSU, there's an additional deduction. GST TDS under section 51 of the CGST Act, at 2% of the taxable value. That's a separate mechanism from income tax TDS. It's credited to your GST cash ledger rather than to your income tax account, and it catches people out because it arrives as a second, unexpected shortfall on the same payment.

Why doesn't the P&L show any of this?

Because the P&L is doing what it's designed to do. Measure performance rather than liquidity.

On the accrual basis your P&L uses, that ₹10,00,000 fee is recorded as revenue in the month you invoiced it. GST never appears in the P&L at all. It's a balance sheet item. A liability you're holding on someone else's behalf. TDS doesn't appear either; it's an advance payment of your income tax, sitting as an asset. So a P&L showing a healthy margin isn't lying to you. It's answering a question you didn't ask.

This is a good reason to learn to read a P&L properly rather than glancing at the bottom line. And an even better reason not to run a business on the P&L alone. The statement that would have warned you is the cash flow statement. The one nobody produces is the one that matters most here: a thirteen-week forward view of money in and money out, by date.

When do I get the TDS money back?

After the financial year ends, you file your return, and the refund is processed. So the money deducted in April is realistically back with you more than a year later.

The mechanism itself is sound. The client deposits the deducted tax against your PAN. It appears in your Form 26AS and Annual Information Statement. You claim it as a credit against your income tax liability when you file. If your actual tax liability is lower than the total TDS deducted across the year, the balance is refunded.

The problem isn't that the money is lost. It's that a growing services business has a permanently growing pile of it. Every month you invoice, another 10% of fees goes into the pile. The pile only unwinds once a year. And if next year's revenue is higher than this year's, the new deductions exceed the old refund. You never get the working capital back. You only ever get last year's smaller version of it back while this year's larger version accumulates.

Growth makes this worse, not better. That's the part founders find genuinely counter-intuitive. It's why revenue growth in a services business can be accompanied by a tightening cash position rather than a loosening one, and why doubling revenue can mean less cash in the bank at the end of the year rather than more.

There are two real levers:

Check your 26AS quarterly, not annually. If a client deducted TDS but never deposited it, or deposited it against the wrong PAN or the wrong quarter, you can't claim the credit. You find this out at filing time, when the client's finance contact has moved on and the correction requires them to revise a TDS return they filed nine months ago. Quarterly checks turn a lost credit into a two-email fix.
Ask your CA whether a lower deduction certificate is appropriate. Section 197 allows an assessing officer to certify deduction at a lower rate where the standard rate would exceed your actual tax liability. It isn't automatic and it isn't for everyone. But for a business whose 194J deductions consistently exceed its final tax bill, it's the difference between financing the government all year and not.

The lever nobody uses: your client's own ITC is on a clock

If a client has claimed input tax credit on your invoice and hasn't paid you within 180 days of the invoice date, they're required to reverse that credit and pay interest on it.

This is Rule 37 of the CGST Rules, read with the second proviso to section 16(2). It's one of the few genuinely asymmetric levers a supplier has. The client has already taken the benefit of the tax you charged them. If they sit on your invoice past 180 days, that benefit is clawed back out of their own GSTR-3B, with interest. And it's a reversal their finance team has to actually process and explain.

Used well, this isn't a threat. It's a reminder that costs the recipient more than it costs you to send. A collections note at day 150 that mentions the approaching 180-day reversal is addressed to a different person than your usual chase email. It lands with the one person in the client organisation who's measured on exactly this.

Two honest caveats though. It only bites if the client has actually claimed the credit, and it does nothing at all about the 30-to-90-day range where most of your pain actually lives. It's a backstop for the genuinely stuck invoice, not a payment-terms strategy. If an account has gone quiet entirely, this is one item inside a longer sequence. I've written the full version of what to do when a client has stopped paying, in the order the steps actually matter.

What size buffer does this business actually need?

Enough to cover one full month of GST plus your fixed monthly costs, held separately from operating cash, before you count any client receipts.

That's the number. Here's how to get to it without a finance function:

Take your average monthly invoiced fees. Not your best month. The median of the last six.
Multiply by 0.18. That's the GST you'll owe on the 20th regardless of what arrives.
Add one month of fixed costs. Salaries, rent, retainers you pay out, software, EMIs. Everything that has a date rather than a decision.
That total is your floor. Cash below it means you're funding the government out of next month's collections. Which works until one large client pays late.

The discipline that makes the floor hold is keeping the GST portion somewhere it can't be casually spent. A separate bank account is the crudest version and it works better than any spreadsheet, for the same behavioural reason a separate account works for building a personal emergency fund. Money you have to make a deliberate decision to move is money you spend far less of. If you want the same arithmetic run for the personal side of your finances, there's an emergency fund calculator on this site that does it.

Does QRMP help my cash flow?

Barely. And this is the most common misunderstanding of the scheme.

QRMP, or Quarterly Return, Monthly Payment, is available to taxpayers with aggregate turnover up to ₹5 crore in the preceding financial year. The name tells you the whole story if you read it carefully. The return becomes quarterly, the payment stays monthly, made through form PMT-06. The government's own QRMP advisory sets out the mechanics.

So QRMP reduces your compliance workload from twelve filings to four. It doesn't defer the cash. If you've been told it will fix your working capital, you've been told wrong. And you should find out what else in that conversation was wrong.

The one genuine cash effect is second-order. Under the fixed-sum method, the monthly payment is calculated from the previous period rather than from current activity, which can smooth a lumpy quarter in either direction. That's a smoothing effect, not a saving.

What to change, in order of effect

If you do nothing else from this post, do these four things, in this order:

1. Show GST as a separate line on every invoice. Cost: one template edit. Effect: TDS is computed on the fee rather than on the GST-inclusive total, permanently.

2. Move invoicing to the start of the month. Cost: a scheduling change. Effect: an invoice raised on the 1st gets 49 days before its GST falls due; one raised on the 30th gets 20. Across a year of invoices this is a structural improvement in average days of float, and it costs nothing. If your contracts allow month-start invoicing and you're invoicing in arrears at month-end out of habit, this is free money.

3. Build the thirteen-week cash view. Cost: an afternoon, once, then twenty minutes a week. Columns are weeks; rows are expected receipts by client, then payroll, then rent, then the 20th-of-month GST payment, then everything else. The value isn't the forecast. The forecast will be wrong. The value is that the week you run out of money becomes visible eight weeks before it arrives, which is enough time to do something other than panic.

4. Reconcile 26AS quarterly. Cost: thirty minutes a quarter. Effect: TDS credits that would otherwise have been silently lost are recovered while the correction is still easy.

And one thing not to do. Don't price the tax gap into your rates as a vague buffer. If your pricing needs to change, change it deliberately and for stated reasons. A margin cushion added to absorb a timing problem hides the timing problem. And the timing problem is fixable while a mispriced service is a much longer conversation. This is the same failure pattern as a retainer that quietly stops being profitable. The cost is real. It's just landing somewhere nobody is looking.

FAQ

How many months of GST buffer should I actually hold in the account?

At least one full month. So 18% of your median monthly invoiced fees, held separately from operating cash, on top of one month of fixed costs. If your receivables regularly slip past 45 days, hold two months of GST instead of one, because that's how long you may need to fund the government before the matching client cash arrives.

What if the client refuses to deduct TDS on the fee alone and deducts on the GST-inclusive total?

Point them to CBDT Circular 23/2017, which explicitly says TDS under section 194J is deducted on the value excluding GST when the tax is shown separately on the invoice. If they still refuse, the extra deduction isn't lost. It goes into your 26AS and you claim it as a refund. But you've financed them interest-free for a year, so it's worth escalating to their tax team once rather than absorbing it every invoice.

Should I switch to QRMP or stay on monthly filing?

Switch if your turnover is under 5 crore and you want fewer filings. But don't switch expecting a cash-flow benefit. QRMP moves the return to quarterly and keeps the payment monthly via PMT-06, so your working capital position is unchanged either way. Pick it for compliance workload, not for liquidity.

What if a client pays in advance before I invoice, does GST still trigger on the invoice date?

No. Under section 13 of the CGST Act, time of supply for services is the earlier of invoice date or payment receipt. If money lands first, that receipt creates the GST liability at the point of receipt, and you should raise the invoice against it promptly so the paper trail matches.

How long does a TDS refund actually take from the end of the financial year?

Realistically anywhere from a few months to more than a year after you file, depending on when your return is picked up and whether 26AS matches your claim. Plan as if the money deducted this year isn't usable until the year after next. And reconcile 26AS every quarter so nothing falls out of the claim at filing time.

Should I invoice on the 1st or the 30th if the contract lets me pick?

The 1st. Without exception. An invoice raised on the 1st gets 49 days before its GST falls due; one raised on the 30th gets 20. Same tax, same client, same amount. The only variable is where in the month the paper is dated, and it's one of the few free improvements to working capital available to a services business.

What if a client sits on my invoice past 180 days, can I really force anything?

You can't force payment. But Rule 37 of the CGST Rules requires the client to reverse the input tax credit they claimed on your invoice, with interest, once 180 days pass unpaid. A polite reminder at day 150 that flags the approaching reversal reaches a different person in their finance team than the usual chase email. And it costs their books more than it costs yours to send.

The part that actually changes the business

None of this is exotic. There's no clever structure here, no aggressive position, nothing your CA will raise an eyebrow at. Four operational habits and one number written on a whiteboard.

But the reason it matters is that a services business almost never dies of unprofitability. It dies of a Tuesday in a month where a large client paid late, payroll was on the 1st, GST was on the 20th, and the founder discovered that the healthy-looking receivables ledger was not the same thing as money. The P&L was right the whole time. It was simply answering a question about performance while the business had a question about liquidity.

Learn the difference between the two calendars. Hold the tax money somewhere separate from operating cash. Give yourself eight weeks of visibility instead of two. The tax system isn't going to change its dates for you, and your clients aren't going to align their payment cycles with the 20th of the month either. So the only variable you get to move is the size of the buffer and the discipline that keeps it intact.

One closing note, because I see this constantly. Verify anything you read here, including this post, against a current primary source or your own CA before acting on it. Rates, thresholds and due dates change, and a confidently-worded answer is not the same thing as a correct one. That applies with particular force to asking a chatbot, which will produce a fluent and entirely plausible GST due date for a regime that changed after its training data ended. I've written separately about why AI tools give wrong answers so convincingly, and tax deadlines are close to the worst possible place to find out.

Good Debt vs Bad Debt: How to Tell the Difference

Prabhash Jha — Fri, 04 Sep 2026 13:31:35 +0000

Home loan good. Education loan good. Credit card bad. Consumer EMI bad. Tidy. It's also how a lot of people end up in trouble while feeling responsible about it.

Look at two positions side by side. A home loan at 8.5% eating 55% of your take-home. A ₹40,000 card balance you clear in two months. The first one is the worse spot to be in, honestly. The label told you nothing. What told you everything was the rate, the size, and whether you can keep paying through a bad quarter.

So drop the good/bad framing. Debt is a price you pay for money you don't have yet. What matters is what the price is, what the money does once you've got it, and what happens to the repayment if your income drops for three months.

This guide walks through the real test, what borrowing actually costs in India, the order to clear debts in, and the narrow set of cases where taking on a loan is genuinely the right call. It's education, not financial advice. No specific products, lenders or securities are recommended anywhere in it.

The real test: rate, purpose, and survivability

Three questions, in this order. A loan has to pass all three.

One: what is the rate, really? Not the advertised rate. The all-in rate after processing fees, insurance premiums bundled into the disbursal, and GST where it applies. Every regulated digital lender in India has to give you a Key Fact Statement showing the annual percentage rate. Ask for it. If a lender drags their feet on producing one, that alone is information.

Two: what does the money do? Money that raises your earning power or buys an appreciating asset is doing work. Money that funds a holiday, a phone upgrade or a wedding is consumption you're paying a premium for. This is where the good/bad framing gets roughly right. It just isn't the deciding factor. A 36% loan for a productive purpose is still usually a bad idea. Almost nothing productive returns 36% reliably.

Three: does the repayment survive a bad month? The one people skip. Lenders assess you on FOIR, fixed obligations to income ratio, and many will happily approve you up to 50–55% of net income. That's their risk appetite, not yours. Their downside is a provision on a balance sheet. Yours is your house.

Run it as a number. Take your monthly take-home, cut it by 30%, and check whether every EMI still clears with something left over. Say your take-home is ₹1,20,000 and total EMIs are ₹48,000. That's 40%, comfortable on paper. Drop the income to ₹84,000 and those same EMIs are 57% of what's coming in, leaving ₹36,000 for rent, food, everything. Not a disaster. But no slack either. At 55% of your original income the same shock leaves you borrowing to pay the borrowing.

The stress test is the whole discipline. Debt doesn't fail because the rate was high. It fails because the payment was due in a month when the money wasn't there. Which is why an emergency fund is a debt-management tool, not a separate topic.

What debt actually costs in India

Rates move, and yours depends on your credit profile, so treat these as typical ranges and check your own sanction letter. The column that matters is the third one. What ₹1,00,000 costs you if you carry it for a year, with compounding at the frequency the lender actually charges.

Borrowing	Typical rate p.a.	Cost of ₹1,00,000 held a year	Foreclosure penalty?
Home loan (floating, repo-linked)	8–9.5%	~₹8,300–9,900	No, barred on floating-rate individual loans
Loan against property	9–12%	~₹9,400–12,700	Usually not if floating
Education loan	9–14%	~₹9,400–14,900	Typically none
Car loan	9–11%	~₹9,400–11,600	Often 3–5% if fixed
Gold loan	9–18%	~₹9,400–19,600	Usually low or nil
Personal loan	11–24%	~₹11,600–26,800	Often 2–5%, lock-in of 6–12 months
Credit card revolve	36–48% (3–3.99% monthly)	~₹52,000–74,000	Not applicable
BNPL / short-tenure app credit	0% headline, often 18–36% real	~₹19,600–42,600	Varies

Two things in that table are worth sitting with.

First, the gap between the top and the bottom row isn't incremental. A home loan at 8.5% costs roughly ₹8,800 a year per lakh. A revolving card balance at 3.5% a month costs somewhere near ₹62,000 per lakh once you account for monthly compounding and the 18% GST that applies to credit card interest. Interest on most other loans is GST-exempt. These are not two versions of the same thing.

Second, foreclosure rules matter more than people expect. Under the RBI's Pre-payment Charges on Loans Directions, 2025, floating-rate loans to individuals for non-business purposes cannot carry a prepayment penalty. Fixed-rate loans and many personal loans can. Before you plan to clear a loan early, check which one you signed.

Worked example: what a card balance really costs

Say you have ₹80,000 sitting on a credit card at 3.5% per month. That's 42% nominal. Interest is charged monthly, so the compounded rate is about 51%. Add 18% GST on the interest. The effective monthly charge is roughly 4.13%. An effective annual cost near 62%.

Now pay only the minimum due, typically 5% of the outstanding. Each month the balance grows 4.13% and then shrinks by 5% of the new figure. So it falls by about 1.1% a month. At that rate it takes roughly 64 months, over five years, just to halve.

Run it out. After those 64 months you'd have paid in roughly ₹1.93 lakh and would still owe about ₹40,000 on an original ₹80,000. You paid more than twice the balance and cleared half of it.

There is no investment that legitimately returns 62% a year, tax-free, with no risk. Clearing that balance is the closest thing to it available to you. That isn't a motivational line. That's the arithmetic.

Two related traps worth naming. Paying part of the bill kills the interest-free period entirely. Interest runs from the transaction date on the full amount, not from the due date on the unpaid bit. And cash withdrawn on a credit card attracts interest from day one plus a cash advance fee, with no grace period at all.

Worked example: what a home loan costs, and what ₹4,339 more a month does

Take ₹50 lakh at 8.5% over 20 years. The EMI works out to about ₹43,391. Over 240 months that's ₹1,04,13,840 paid, of which ₹54,13,840 is interest. You pay more in interest than you borrowed.

Now raise the EMI by 10%. That's ₹4,339 more a month, to ₹47,730. The loan closes in about 192 months instead of 240. Total paid: ₹91,64,160. You save roughly ₹12.5 lakh in interest and finish four years early, for the price of one modest annual increment redirected.

That's the single highest-leverage move available on a long loan. No product needed. No advisor. No market view. It works because interest is charged on the outstanding balance, so money paid early removes interest from every remaining month.

The tax angle, honestly. Under the old tax regime, Section 24(b) lets you deduct up to ₹2 lakh of home loan interest on a self-occupied property. In year one of that ₹50 lakh loan the interest is about ₹4.21 lakh. So the deduction covers less than half of it. At a 31.2% marginal rate the saving is ₹62,400, which pulls the effective cost from 8.5% down to roughly 7.2%. Under the new regime, which is now the default, that deduction isn't available on a self-occupied home at all. The rate is simply the rate. Anyone still repeating "take a home loan for the tax benefit" is quoting a rulebook most borrowers no longer sit under. Check which regime you're actually filing in before you build a plan on it.

Education loans work similarly. Section 80E allows a deduction on the interest with no upper cap for up to eight years, but only under the old regime.

The order to repay in

Two well-known approaches. The choice between them isn't really about maths.

Approach	Method	Cost	Use it when
Avalanche	Clear the highest interest rate first	Cheapest in absolute terms	You'll stick with it without visible wins
Snowball	Clear the smallest balance first	Costs more in interest	You've stopped and restarted before and need momentum

Avalanche is mathematically correct. Snowball wins where behaviour is the binding constraint, which for many people it is. The worst option is neither. Paying a bit extra on everything at once. That finishes nothing.

A practical priority order:

Anything above 20%. Revolving card balances, overdue BNPL, informal borrowing. Treat as an emergency and clear before you do anything else with spare money.
Personal loans in the 14–24% band. Expensive, unsecured, no asset behind them.
Hold a one-month EMI buffer in cash. Before you accelerate anything further, make sure a bad month doesn't push you back onto the card you just cleared.
Secured mid-rate debt (9–12%). Car, gold, loan against property. Gold loans deserve attention despite the modest rate. The tenure is short and if gold prices fall the lender can demand a top-up or auction the collateral.
Sub-10% long-tenure debt. Home and education loans. Lowest priority. If they're comfortable and you have deductions running, there's a reasonable argument for not rushing them.

Where "good debt" quietly goes bad

The home loan that's too big. The loan is fine. The ratio isn't. Once the EMI passes about 40% of take-home, every other financial decision gets made under duress. And a home is illiquid. You can't sell 15% of it to cover a bad quarter.

The education loan taken without a payback calculation. Compare the loan's total cost against the realistic salary difference the qualification produces, and how many years it takes to close the gap. If the honest answer is nine years, it's an expensive purchase, not an investment. The label "education" doesn't do the work.

The business loan used to cover a cash gap. Borrowing to buy equipment that produces output is one thing. Borrowing because customers pay in 90 days and salaries are due in 30 is a working capital problem. A term loan is the wrong instrument for it. You're solving timing with a five-year commitment. This is exactly the cash flow versus profit distinction that closes profitable businesses. One genuine advantage worth noting: business loan interest is a deductible expense, so at a 25% tax rate a 12% loan costs about 9% after tax.

BNPL, no-cost EMI, and loans that don't feel like loans

"No-cost EMI" isn't free credit. The interest is either embedded in a discount you'd otherwise have received, or it reappears as a processing fee. It's a loan. It appears on your credit report. And it consumes borrowing capacity you may want later.

BNPL is the same trade with worse consequences on default. RBI's digital lending rules require disbursal and repayment to move directly between your bank account and the regulated lender, and require that Key Fact Statement with the APR. If money is routing through a third party, or nobody will show you an APR, that's a reason to stop.

The real damage from small BNPL amounts is rarely the interest. It's the late fees and the credit report entry. A missed payment reported to the bureaus follows you for years. A loan closed as "settled" rather than "closed" is worse than paying in full. It flags to every future lender that you didn't repay the agreed amount. Check your report. You're entitled to one free full report a year from each bureau.

One more thing to watch on the paperwork. Lenders often bundle a single-premium loan protection policy into the sanctioned amount. You then pay interest on the premium for the whole tenure, which raises the effective rate above the number on the front page. Sometimes the cover is worth having. The point is to price it separately and decide, not absorb it. And note that anyone earning commission for referring an insurance product in India needs IRDAI registration. That includes a website that links to one.

When borrowing is genuinely correct

Narrow list. A loan makes sense when the rate is low, the money buys something that produces income or holds value, the repayment survives the stress test, and you have a plan for the loan's full tenure. Not just its first year.

So: a home you can comfortably afford. A qualification with a calculable payback. Equipment or inventory that generates more than it costs. Occasionally a genuine emergency at a secured rate, when the alternative is breaking a long-term asset at the wrong moment.

What it doesn't cover: consumption at any rate above roughly 12%. And borrowing against an expectation of future income you haven't yet earned.

FAQs

Is all credit card debt bad?

The card isn't the problem. Revolving a balance is. Paid in full each month, a card is free short-term credit and builds your record. Carried at 3.5% a month with GST on top, the effective cost is around 62% a year. Also remember that a partial payment removes the interest-free period entirely, so interest runs from the transaction date.

Should I invest or pay off debt first?

Compare the loan rate to what you could realistically earn after tax. Clearing a 40%-plus card balance is a guaranteed, tax-free return nothing else matches, so that comes first. An 8.5% home loan is a closer call and can reasonably run alongside investing. This is a framework, not a recommendation of any specific investment.

How much EMI is too much?

As a working ceiling: total EMIs under 40% of take-home, with home loan EMI under 35%. Lenders will approve you well past that. FOIR limits of 50–55% are common. But that's their tolerance, not your safety margin. The real test is whether the payments still clear if your income falls 30% for three months.

Is a home loan good debt in India?

Usually the cheapest borrowing you'll get, at 8–9.5% floating and no foreclosure penalty on individual floating-rate loans. But under the new tax regime the Section 24(b) deduction on a self-occupied property is gone, so the tax argument no longer holds for most borrowers. Judge it on affordability and on the property, not the label.

Does prepaying a home loan actually help?

Substantially, and earlier is better. On ₹50 lakh at 8.5% over 20 years, raising the EMI by 10%, about ₹4,339 a month, closes the loan in 16 years instead of 20 and saves roughly ₹12.5 lakh in interest. Confirm your loan is floating-rate first. Fixed-rate loans can carry a prepayment charge.

Is no-cost EMI really free?

No. The interest is either built into a discount you forgo or charged as a processing fee. It's a formal loan that appears on your credit report and uses up borrowing capacity. Ask for the Key Fact Statement with the APR. Regulated lenders must provide one. Compare the EMI price against the outright cash price before deciding.

Key takeaways

The good/bad label decides nothing. The interest rate, what the money does, and whether the EMI survives a 30% income drop decide everything.
A revolving credit card balance at 3.5% a month costs roughly 62% a year once monthly compounding and the 18% GST on card interest are counted. Clearing it beats any investment available to you.
Paying only the minimum due on ₹80,000 takes about 64 months just to halve the balance, by which point you've paid in around ₹1.93 lakh.
On a ₹50 lakh home loan at 8.5%, a 10% higher EMI ends it four years early and saves about ₹12.5 lakh in interest.
Under the new tax regime there is no Section 24(b) deduction on a self-occupied home, so the old "take a loan for the tax benefit" argument no longer applies to most borrowers.
Keep total EMIs under 40% of take-home and hold a one-month EMI buffer in cash. Lenders will approve you far past that, but their risk appetite isn't yours.

Deciding What an AI Agent Is Allowed to Touch: the Permissions Boundary That Survives a Bad Day

Prabhash Jha — Thu, 03 Sep 2026 13:34:02 +0000

The real question about an AI agent isn't "can we make it do the thing?" It's "what does the recovery look like the day it does the thing wrong?"

Most writing on AI agent permissions splits into two piles. One is identity-vendor content, all Okta-shaped diagrams and scope tables and OWASP taxonomies. The other is enterprise-security framing, SOC 2 evidence and EU AI Act cross-references. Both are useful somewhere. Neither is what a small operator needs on the day they hand a machine the keys to their git repository or their live production site. What they need is a decision framework that answers one question honestly per action the agent could take: what's the worst plausible outcome on a bad day, and what does the recovery from that look like?

This post is that framework. I wrote it from having actually handed agents write access to this website. Everything below is either something I do, or something where I've decided the recovery cost is too high to justify the convenience. It isn't a security-vendor pitch. It doesn't sell you a policy engine. It's the shape of the decision itself, written for the person who's going to eat the consequences.

The framing is deliberately small-operator. A team of one, two, or five, running unattended jobs against their own infrastructure. The reason that framing matters is that most permissions writing assumes a security team behind you. There isn't one. You are the security team. The framework has to fit in your head, not in a compliance binder.

Why "least privilege" is the wrong first question

The industry-standard framing is least privilege: give an agent only what it strictly needs. Fine principle. Very late one. It answers "how much access?" without first answering "which class of action is safe to grant any level of access to?"

The order that actually helps a small operator is:

Classify the action by blast radius. Not by whether it can be done. Anything an API supports can be done. Classify by what happens if it goes wrong and cannot be undone. This is where permissions decisions should start and where they almost never do.
Decide whether the action belongs to a class you'll let agents perform at all. Some classes just don't, no matter how tightly scoped the token is.
Only then, for the actions that pass the first two gates, apply least privilege to the specific scopes and tokens.

Skip steps 1 and 2 and you end up with an agent that "just needs delete permission for that one folder". A bad day later, it has removed something you can't get back. The scope was narrow. The class was wrong. Nobody caught the difference because the vocabulary the industry gave you doesn't distinguish between them.

The blast-radius taxonomy that decides everything else

There are essentially four classes of action. The class is what determines whether an agent should ever touch the action. Not the API's scope list.

Class 1. Reversible by the agent itself, within seconds. The agent edits a draft. The agent runs a query and inspects the result. The agent generates content into a file the agent also owns. Recovery from a bad Class 1 action is that the agent tries again. There's no meaningful blast radius because the action never left the agent's own sandbox. Grant these freely. Not doing so is what makes AI agents feel useless.

Class 2. Reversible by a human in minutes. The agent commits to git. The agent pushes to a branch. The agent uploads a draft to a CMS. The agent enqueues an email in a "review before send" queue. Recovery is that a human notices, reverts, or reviews. The blast radius is real but bounded. A bad commit is a bad commit, and git revert is a real answer that costs about three minutes. Grant these, with the audit trail that lets the human notice.

Class 3. Reversible in hours or days, at real cost. The agent posts publicly. The agent sends email to a real list. The agent modifies a live production database. Recovery is that a human notices AND takes an action that itself has consequences. Deleting a public post. Sending a correction email. Rolling back a migration. Grant these only if the specific action's blast radius is small AND the audit trail lets you catch the bad case within minutes, not hours.

Class 4. Irreversible, or reversible only at business-changing cost. The agent deletes a customer record. The agent moves money. The agent changes account credentials. The agent posts something that gets picked up by the wire before the correction lands. Recovery is that you spend a week apologising or paying, and the trust cost is permanent. Do not grant these. Ever. No matter how competent the agent, no matter how narrowly scoped the token. The convenience of automating a Class 4 action is never worth the cost of getting it wrong once.

The lines between the classes aren't fuzzy in practice. What makes them feel fuzzy is that the industry sells you tools that can technically perform every action, and lets you decide which. The taxonomy above is the pre-tool decision. Before you write a single line of code that could touch a Class 4 action, ask yourself whether the action belongs in the code at all. If the answer is no, the tooling question never comes up. If the answer is yes, you've made a decision you can be held to on a bad day. That distinction is the whole point.

Notice what the classes are graded on. Time to recover, and cost of recovery. Not "how important is the system" or "how sensitive is the data". Sensitivity matters for other reasons. It isn't what changes the class of the action.

One case sits awkwardly across the taxonomy and is worth calling out, because it is the one most people ship first: an agent that only talks. No tools, no writes, nothing that looks like a Class 3 action anywhere in it. The trap is that a sentence said to a customer in your name can itself be a commitment you are held to, which moves it up the scale without ever touching an API. What actually breaks when you put a language model in a customer-facing flow works through that case specifically — the failure modes, what they cost, and why "it can only reply" is not the reassurance it sounds like.

The specific things I let my agents do

For concreteness, here's the exact scope I grant on this site, mapped to the classes above. This isn't a template. Your operation is different. It's a worked example of the framework.

Class 1, granted: anything that reads. Read repo files. Read the sitemap. Read the deploy logs. Read the digest. Read the worklist. Read the backlog. The publishing agent reads everything it can, all the time, and there's no recovery cost because reads don't change anything.

Class 1, granted: anything that writes to files owned by the agent's session and inspected before commit. Draft posts. Generated covers. Generated OG cards. Updated backlog metadata. The bad case is a bad draft. The recovery is "delete the file, try again".

Class 2, granted: commits to main when the build gate passes. This is the important one. I explicitly grant the agent the ability to write to the branch that Cloudflare Pages deploys, without a human review step, on the specific condition that npm run build returns zero. The build gate is what makes this a Class 2 action rather than a Class 3 one. A broken deploy gets caught by check-posts, check-links and check-search before it lands. A bad commit that passes the gates is still git revert-able within minutes.

Class 2, granted: pushes to origin/main. Same reasoning. The deploy pipeline has its own build gate on the far side, and Cloudflare Pages retains previous deployments for rollback. A single click undoes any single deploy.

Class 2, granted: GitHub API calls to create commits, open issues, update issue state on the worklist. All auditable in the repo history. All undoable by hand.

Class 3, considered and NOT granted: publishing to LinkedIn or any social platform under my name. The recovery cost of a bad post on LinkedIn is a deleted post that everyone who saw it also screenshot-shared, plus a reputational tax. The convenience of automating "post the article as a LinkedIn update" isn't worth that. This is the specific reason the site's AUTOMATION-PLAN document says explicitly "It does not touch LinkedIn or email". The class was considered and rejected, not just left undone.

Class 3, considered and NOT granted: sending email to the newsletter list. Same reasoning. The recovery from a bad email to a real list is a correction email that people also read as "these people can't keep their systems straight", plus unsubscribes that are permanent. Newsletters ship when a human clicks send.

Class 4, hard-blocked: deletion of any published post. Deletion of any subscriber record. Changes to _headers, _redirects, functions/, or package.json dependencies without an explicit re-review. The agent can PROPOSE these in a run log entry. The commit that lands them has to be one I read, not one the agent authored. It's a real friction on the agent. It's a friction I want.

Class 4, hard-blocked: anything to do with Cloudflare account settings, DNS, or Access rules. The account credentials aren't in the agent's environment at all. If they need changing, I change them.

The two questions to ask before granting an agent any action

Before you extend an agent's permissions to include a new capability, force yourself to answer both of these in one paragraph each. Not a checklist. Actual sentences.

Question 1: What's the worst plausible outcome of this action going wrong, and how long does the recovery take? Not the average outcome. The worst plausible one. If the agent could theoretically do the action ten thousand times before you notice, use that number, not one. "The agent could send one wrong email" is a Class 3 action. "The agent could loop and send the same wrong email to a thousand recipients before anything catches it" is a Class 4 action. Autonomy is what changes the class.

Question 2: What class of person is the recovery going to require, and are they going to be available on a bad day? A recovery that requires you personally to be reachable within 15 minutes for the recovery to still be a Class 2 action isn't really Class 2. It's Class 3 with a lucky timing assumption. Be honest about who else has the credentials to fix a bad case and how reliably they can be reached.

If you can't answer both questions to your own satisfaction in one paragraph each, the answer to "should I grant the action?" is no.

The three failure modes I have actually seen

Failure mode 1: "This is only for the demo". An agent gets a broader-than-needed permission during development because it's faster to grant everything and narrow later. Then the demo works. The demo becomes staging. Staging becomes production. Nobody remembers to narrow the permissions until the day the agent hits a case its narrowed-later self would have refused. Grant narrow from the first line, even during development. It takes ten minutes at the start. It's impossible after go-live because real work now depends on it.

Failure mode 2: "The scope doesn't include destructive actions". GitHub's repo scope, for example, includes deleting branches, force-pushing, deleting files. It looks like a read-write scope. It's actually a "do anything to this repository" scope. Read the scope's actual permissions before granting it. Not the scope's name. Most OAuth scopes bundle actions the caller doesn't intend to allow. The bundling is where the blast radius lives.

Failure mode 3: "We'll add monitoring later". Any Class 2 or Class 3 permission granted without the audit trail that lets you notice a bad action is silently a Class 4 action. The recovery clock doesn't start until somebody notices. If nobody notices, there's no recovery. Just accumulated damage. The audit trail defines the class as much as the action does. This is the whole argument in what to log when an AI agent acts on your behalf. Logging isn't documentation. It's the containment measure.

The "does the human belong here?" question

Once you've classified an action, one specific decision separates well-run automation from theatre. Do you put a human in the loop?

The industry default is yes, on anything Class 3 or above. That's the wrong answer, and it's the wrong answer for the same reason "just add a review step" fails everywhere else. A human who has to approve two hundred things a week isn't a control. They're a rubber stamp with a chair. The human check is only real if the human can actually tell good output from bad, and if the volume is low enough that they can pay attention.

Two rules that work in practice:

A human check is only a control when the human can refuse. If refusing is expensive (blocks the release, delays the client, means a difficult conversation), the human won't refuse routinely, so the check is theatre. If the reviewer's job is to be right, not to be popular, and refusing is genuinely free, the check works.
A human check on 20 items a week is a control; a human check on 200 items a week is not. Below a threshold, humans pay attention. Above it, they stop reading. The threshold varies by task but it's much lower than most systems assume. I use "if I would open every one of these individually, it's under the threshold; if I skim, it's not".

Where a human check is theatre, the honest answer isn't "add a better checklist". It's either: automate the specific decision entirely (the check isn't adding value; remove it), or reduce the volume until a human can actually make it (the check is important; the volume is the problem, not the check). Automation with a rubber-stamp step is worse than automation without one. The presence of the step deflects the blame away from the design that made the volume too high.

The specific artefacts a permissions decision needs

For every action class you grant to an agent, three specific artefacts have to exist. Miss any of the three and the permission hasn't been granted responsibly. It's been granted optimistically.

Artefact 1. The audit trail from outside the agent. Not the agent's own log of what it did. An external record. A git history. An API response log. An email queue with message IDs. Something that survives the agent dying, the process being killed, the machine losing power. The action's real evidence lives outside the actor. See the full argument in what to log when an AI agent acts on your behalf. Short version: the agent is the least reliable witness to its own actions.

Artefact 2. The recovery procedure, written down. A one-paragraph document per action class that says: "If action X fires wrong, do Y and Z in that order, and expect the recovery to complete in W minutes." If the procedure doesn't exist, the action isn't really recoverable. It's unrecovered but not yet detected. Write the procedure before the first time you grant the permission, not after the first time you need it.

Artefact 3. The kill switch. A single named place where you can revoke the agent's ability to perform this action within seconds. For me, on this site, the kill switches are: rotating the GitHub OAuth token (kills all write access), rotating the Cloudflare API token (kills the deploy verification), and setting the daily-run cron to disabled (kills the trigger). All three are documented. All three are one command each. All three are things I've used at least once.

Without a real kill switch, the permission isn't a permission. It's a fait accompli. Grant nothing you can't revoke in under a minute.

The single hardest one: standing versus just-in-time permissions

The most common mistake in agent permissions is granting standing access to a Class 3 action so the agent can "handle it when needed". Standing permissions are the reason Class 3 actions turn into Class 4 problems. An idle or compromised agent sitting on standing privilege isn't idle. It's exposed. This is why Microsoft's own guidance and industry security research on AI agent identities both converge on the same recommendation: grant elevated permissions only for the moment they're needed, with a short TTL and automatic revocation.

The just-in-time model looks like this. The agent asks for the permission when it needs it. It gets a token with a lifetime measured in minutes not months. The token expires whether the agent needed to use it or not. If the agent doesn't need the permission for the next hour, no exposure exists. If the agent is compromised in that hour, the attacker inherits a token that expires before they can pivot.

More work to set up. Also the difference between a permission model that survives a bad day and one that turns a bad day into a bad quarter. If you're only going to build one non-obvious piece of infrastructure for your agent, this is it.

What to do next Tuesday

If you already run agents on your own systems, do the following two exercises this week.

Exercise 1. Write down every action your agent can currently perform. Not the tool names. The actions. "Commit to git." "Send an HTTP POST to my analytics endpoint." "Read a database row." "Write a database row." For each one, mark the class from the taxonomy above.

Exercise 2. For every Class 3 or Class 4 action on the list, answer the three-artefact test: audit trail from outside, written recovery procedure, kill switch. Any action that fails on any of the three, either downgrade the action (revoke the permission until the artefacts exist) or take the time to build the artefact before the next run.

If the exercise takes an hour, you have a small operation and this is worth an hour. If it takes a week, you have a larger operation and the delay of doing it now is smaller than the cost of the first bad day you haven't prepared for.

Where this fits with the rest of the work

The permissions boundary is the containment side of running unattended agents. The audit trail is the diagnostic side, in what to log when an AI agent acts on your behalf. The log tells you why the agent did what it did. The permissions limit what the agent could have done at all. Together they're what make an autonomous system operable.

If your problem is upstream, if you haven't yet decided which tasks to hand to an agent at all, the framing in how to automate your work with AI is the earlier question. And if your specific failure mode is the workflow that runs green and does nothing, a silent Class 2 action pretending to be a completed one, the diagnosis is in the automation failure nobody catches.

FAQ

How long should a just-in-time agent token live?

Long enough for the specific action, and no longer. For a commit-and-push cycle that's minutes. For a batch job that runs in a loop, it's the length of the loop plus a small buffer. If you can't state the number in minutes, the token is standing access wearing a costume.

Should I put a human review step on every Class 3 action, or automate it fully?

Neither by default. If the reviewer can genuinely refuse without punishment and the volume is low enough that they read every item, keep the review. Otherwise pick one. Automate the decision entirely, or cut the volume until the review is real. A rubber-stamp step is worse than no step because it launders responsibility.

What if my agent needs a Class 4 permission for one specific edge case?

Then that edge case isn't automated. The agent can prepare the action, log the request, and page a human. The commit, the send, or the delete is done by the human. The convenience of automating one Class 4 case is never worth building the plumbing that could execute all of them.

How many actions is too many for one agent token?

Count the classes, not the actions. One token holding any mix of Class 1 and Class 2 actions is fine. A single token that spans Class 2 and Class 3 is where blast radius creeps in silently, because the audit expectations for the two are different. Split tokens along class boundaries, not along feature boundaries.

Do I need a kill switch if the agent only runs on a schedule?

Yes. A scheduled agent that has already fired and is mid-loop is exactly when you need to revoke access, and "wait for the next window" isn't a kill switch. If rotating the token or disabling the trigger isn't a single documented command you've run at least once, you don't have one.

Should the agent write its own audit log, or should I rely on external systems?

External systems. The agent is the least reliable witness to its own actions. A crashed, killed, or compromised agent won't finish writing its log. Git history, API response logs, and message queues survive the agent dying. The agent's own log is a convenience, not evidence.

What is the fastest way to audit permissions I have already granted?

List every action the agent can currently perform in plain English. Mark each with a class from the taxonomy. For every Class 3 or Class 4 line check the three artefacts: external audit trail, written recovery procedure, one-command kill switch. Any row missing an artefact gets revoked until the artefact exists. An hour of this work is cheaper than the first bad day you haven't prepared for.

The one-sentence version

Classify every agent action by blast radius before you classify it by scope. Refuse Class 4 actions entirely no matter how tightly scoped the token. Require three artefacts (external audit trail, written recovery procedure, one-command kill switch) for every Class 2 and Class 3 permission you grant. Never grant standing access to a Class 3 action when just-in-time will do. The permissions model that survives a bad day is the one designed around the bad day, not the good one.

Cash Flow vs Profit: The Difference That Sinks Most Small Businesses

Prabhash Jha — Thu, 03 Sep 2026 13:34:02 +0000

Most people assume businesses fail because they lose money. They don't. Losing money is a slow death, and it's usually visible for months before it lands. The fast death, the one that genuinely surprises owners, happens to businesses that are profitable.

A business can be profitable on paper and still go broke. Happens all the time. Profit and cash flow aren't the same thing, and it's cash flow that keeps the lights on.

Here's the part that trips people up. The two numbers don't just differ in size. They differ in timing. Profit answers "did this month's work create value?" Cash answers "can I pay Friday's salaries?" A growing, well-priced, well-run business can answer yes to the first question every month and no to the second one in month four. That isn't bad luck. It isn't bad management. It's arithmetic, and it's predictable if you know where to look.

This post shows you the arithmetic. A month-by-month example where a business books ₹30,00,000 of profit in six months and ends the period overdrawn. Then the reconciliation showing exactly where the money went. Then the fixes that put it back.

Profit and cash flow answer two different questions

Profit is an accounting figure. Revenue minus expenses, matched to the period the work was done in. Deliver goods on 30 June, that sale is June revenue. It doesn't matter if the customer pays in July, October, or never. If the layout of that statement is unfamiliar, start with how to read a profit and loss statement without an accounting degree and come back.

Cash flow is the actual money moving in and out of your bank account, and when. Promises don't count. An invoice isn't cash. A signed purchase order isn't cash. A payment that has cleared is cash.

The gap between them lives in the balance sheet. Every rupee sitting in receivables (invoiced, not collected) or in inventory (bought, not sold) is profit you have earned and cash you don't have. Every rupee sitting in payables (received, not paid) is the reverse. Cash you're holding that isn't yours.

Transaction	Effect on profit	Effect on cash
You invoice a customer ₹5,00,000 on 60-day terms	+₹5,00,000 now	Nothing for 60 days
Customer pays that invoice	Nothing	+₹5,00,000
You buy ₹3,00,000 of stock for next quarter	Nothing until sold	−₹3,00,000 now
You take a ₹10,00,000 loan	Nothing	+₹10,00,000
You repay ₹1,00,000 of loan principal	Nothing	−₹1,00,000
Monthly depreciation on equipment	−₹40,000	Nothing
You pay a supplier from last month	Nothing	−the amount
Customer pays a 30% deposit before you deliver	Nothing yet	+30% now

Read that table twice. Half the rows move one number and not the other. Loan principal, deposits, stock purchases and collections are pure cash events. Your P&L will never show them to you. Depreciation is a pure profit event that never touches your bank. Manage the business off the P&L alone, and you're blind to four of those eight rows.

Why profitable businesses still run out of money

Four causes do most of the damage.

Late-paying customers. The sale is booked as profit the day you invoice. The cash lands 60 or 90 days later. If you chase it.
Money tied up in inventory. Profit sitting on a shelf isn't cash in the bank. You've already paid for it.
Growing too fast. Growth eats cash. You pay for more stock, more staff and more ad spend before the revenue from that growth catches up.
Big lumpy bills. GST, advance tax, annual insurance, bonuses and supplier settlements land on their own schedule. Not yours.

The first three are the same mechanism seen from different angles. Your money goes out before it comes in. The bigger you get, the more of your money sits stuck in that gap at any moment. Growth doesn't cause the problem. Growth multiplies a problem that was already sitting in your terms.

Worked example: ₹30 lakh of profit, ₹7 lakh overdrawn

Take a small B2B supplier. Gross margin is a healthy 30%. Fixed overheads are ₹4,00,000 a month, paid in the month. Customers are corporates on 90-day terms. Suppliers give 30 days. Opening bank balance is ₹6,00,000. The business has been running steadily at ₹16,00,000 a month before this period.

Now it grows. ₹20 lakh, ₹24 lakh, ₹28 lakh, ₹32 lakh, ₹36 lakh, ₹40 lakh over six months. Doubling in half a year. Every month is profitable. Profit rises every month.

All figures in ₹ lakh (1 lakh = ₹1,00,000).

Month	Revenue	Net profit	Cumulative profit	Cash in	Cash out	Closing bank
1	20.0	2.0	2.0	16.0	15.2	6.8
2	24.0	3.2	5.2	16.0	18.0	4.8
3	28.0	4.4	9.6	16.0	20.8	0.0
4	32.0	5.6	15.2	20.0	23.6	−3.6
5	36.0	6.8	22.0	24.0	26.4	−6.0
6	40.0	8.0	30.0	28.0	29.2	−7.2

Compare two columns. Cumulative profit, and closing bank. Profit climbs steadily to ₹30,00,000. The bank account drains to zero by month three and is ₹7,20,000 overdrawn by month six.

The mechanics. Cash in during month 4 is ₹20 lakh, because that's month 1's revenue arriving on 90-day terms. Cash out in month 4 is month 3's cost of goods (₹19.6 lakh, paid on 30-day terms) plus ₹4 lakh of overheads. You're paying for month 3's volume out of month 1's collections. While you're growing, month 3 is always bigger than month 1. So the gap widens every single month.

Notice the sequencing. The business doesn't look sick in month 1 or 2. Month 3 is the first month it can't absorb a surprise, and month 3 is also the month the owner is most likely to be celebrating, because the P&L has never looked better. Month 4 is where salaries get funded by an overdraft.

The trap is what the P&L is telling you at that moment. Revenue climbing. Margin holding. Every conversation with the accountant ends with "good month". Meanwhile the bank statement is doing something completely different. Two documents, same business, two stories.

Where the ₹43 lakh went

Six months of profit came to ₹30,00,000. Over the same six months, cash fell by ₹13,20,000 (from ₹6,00,000 to −₹7,20,000). The difference is ₹43,20,000. It didn't vanish. It moved into working capital.

Receivables rose ₹60,00,000. At the start you were owed three months of ₹16 lakh sales, so ₹48,00,000. At the end you're owed the last three months of a much bigger business: ₹32 + ₹36 + ₹40 lakh = ₹1,08,00,000. That ₹60 lakh increase is money you earned and handed to your customers as free credit.
Payables rose ₹16,80,000. You owe suppliers one month of cost of goods. That was ₹11,20,000 at the start and ₹28,00,000 at the end. This one works in your favour. It's cash your suppliers are lending you.

Net working capital absorbed ₹60,00,000 − ₹16,80,000 = ₹43,20,000. And ₹30,00,000 of profit minus ₹43,20,000 of working capital equals −₹13,20,000 of cash. The reconciliation is exact.

The whole lesson in one line. When you grow, the gap between what customers owe you and what you owe suppliers grows with you, and you fund the difference out of your own pocket.

Growth is a cash expense, budget for it

There's a simple way to size how much cash your growth will eat before you commit to it. Measure your cash conversion cycle.

CCC = days sales outstanding + days inventory outstanding − days payables outstanding

In the example, 90 days of receivables, roughly 0 days of inventory (buy-to-order), 30 days of payables. CCC = 60 days. Which means every rupee of cost you incur leaves your bank 60 days before the matching rupee of revenue arrives.

The working capital you must fund at any moment is approximately:

(DSO × daily sales) + (DIO × daily cost of goods) − (DPO × daily cost of goods)

At the end of month 6, receivables hold the last three months of invoices. ₹32 + ₹36 + ₹40 lakh = ₹1,08,00,000. Against ₹28,00,000 of cost your suppliers are funding. That's about ₹80,00,000 you have to find yourself.

And it gets worse before it gets better. If sales simply hold flat at the month-6 run rate of ₹40,00,000, receivables settle at a steady 90 days of sales. ₹1,20,00,000 against ₹28,00,000 of payables. A gap of ₹92,00,000. The ramp hasn't finished washing through. Stopping growth doesn't release the cash. It only stops the hole getting deeper. The cash comes back only when you shrink, or when you change the terms.

Run this calculation before you take the bigger order. If a new contract adds ₹10,00,000 a month of revenue at a 60-day CCC, you need roughly two months of its cost sitting in your bank before you sign it. A contract you can't fund isn't an opportunity. It's a liability with a nice logo on it.

The other way to look at CCC is as a tax on growth. Every extra rupee of sales carries a cash charge you pay upfront and recover months later. Cheap when you have the buffer. Ruinous when you don't. The number of small businesses that grew themselves into insolvency doesn't get talked about because nobody wants to admit they died winning.

The Indian timing traps that are easy to miss

Three of these hit small Indian businesses hard. None of them appear in a P&L view of the world.

GST is payable on invoices, not collections. For most registered businesses, liability arises at the time of supply. Broadly, when you raise the invoice. You remit the tax with your monthly GSTR-3B (or under the QRMP scheme if you qualify by turnover) while the customer is still sitting on your invoice. On ₹40,00,000 of monthly sales at 18%, that's ₹7,20,000 of output tax leaving your account, offset by input credits, on money you haven't received. Confirm your own filing frequency and due dates with your CA.

TDS shrinks the cheque. Corporate customers deduct tax at source. Commonly 2% under section 194C for contract work or 10% under 194J for professional services. You booked the full invoice as revenue. The bank receives less. You recover the difference as credit when you file your return, which may be many months away. Plan cash on the net figure, not the invoice figure.

Advance tax is lumpy. Instalments fall due in June, September, December and March. Forecast on a smooth monthly average and you'll be short in exactly those four months.

The first two traps are severe enough in a services business to deserve their own treatment. The two calendars involved, the invoice date the tax runs on and the payment date your client runs on, are what turn a profitable retainer book into an empty bank account. I've written the full version separately. Why GST and TDS both hit your cash before the client does, including the invoice-layout detail that decides whether TDS is deducted on your fee or on your GST-inclusive total, and the buffer arithmetic that tells you how much cash this business actually needs to hold.

One rule works in your favour, though. Under the MSMED Act, buyers must pay registered micro and small enterprises within the agreed date, capped at 45 days, and delayed payment attracts compound interest at three times the RBI-notified bank rate. Section 43B(h) of the Income Tax Act reinforces this by disallowing the buyer's deduction if they pay a micro or small supplier late. If you're Udyam-registered, say so on your invoices and in your terms. Larger buyers with competent finance teams are aware of the consequences, and it changes how your invoice is queued. Delayed payments can be escalated through the government's Samadhaan portal. Take specific advice from your CA on your own registration status. When an invoice actually does go bad rather than merely slow, there is a sequence to follow and an order to follow it in. Including when to stop work, and the one thing you should never withhold.

What actually fixes it

Five levers, in order of how quickly they work.

Watch cash, not just profit. Know what's in the bank and what's due when. A weekly ten-minute review beats a beautiful monthly P&L that arrives three weeks late.
Get paid faster. Invoice the day you deliver, not at month end. Ask for deposits. Chase late payers without guilt. You aren't asking for a favour. You're asking for money you've already earned.
Keep a cash buffer. A few months of fixed costs turns a scary month into a manageable one. The business version of the same logic behind a personal emergency fund.
Slow the money going out. Negotiate supplier terms, avoid over-ordering stock, and stop pre-paying for annual things you could pay monthly.
Forecast. A month-by-month cash forecast warns you of a squeeze while you can still do something about it.

Here's what those levers are worth, applied to the same business.

Lever	What changes	Cash effect over the six months
30% deposit on order, balance at 90 days	Part of every sale collects immediately	+₹32,40,000
Cut terms from 90 to 60 days	One month of sales released	Roughly one month's revenue, once
Extend supplier terms 30 → 45 days	Half a month of cost funded by suppliers	Roughly ₹14,00,000, held
Raise price 5%, volume flat	Margin 30% → 33.3%	+₹9,00,000 of profit and cash
Delay one month of growth	Smaller receivables balance	Reduces the gap, does not close it

Take the deposit lever. It's the strongest and the most underused. Charge 30% on order and collect the remaining 70% on the usual 90 days. Month 1 receipts become ₹16.0 lakh (the old invoices still arriving) plus ₹6.0 lakh of new deposits = ₹22.0 lakh. Run that through all six months and the closing bank balance is +₹25,20,000 instead of −₹7,20,000. A ₹32,40,000 swing, with identical revenue, identical costs and identical profit.

Be honest about one thing. Part of that swing is a one-off transition benefit. In the first three months you collect both the old full invoices and the new deposits. But the structural gain is permanent. Once 30% of every sale arrives on day zero, your effective DSO drops from 90 days to 63. And it stays there.

Notice what this lever did not require. No new customers. No price increase. No cost cutting. No bank. Just a change to when money moves. Which is why cash is the cheapest problem in a business to fix. And the most commonly ignored.

Build the forecast: 13 weeks, one sheet

Monthly forecasts hide the problem. Salaries, GST and supplier runs cluster around particular dates. Work in weeks.

Down the rows: opening balance, then cash in (by named customer and expected date, not by average), then cash out (payroll, suppliers by name, rent, GST, loan EMIs, ad spend), then net movement, then closing balance. Across the columns: the next 13 weeks.

Three rules make it useful rather than decorative.

Use expected dates, not invoice dates. If a customer has paid on day 75 for the last four invoices, forecast day 75. Their stated terms are aspiration. Their history is data.

Update it weekly and keep last week's version. The gap between forecast and actual is the number that teaches you something. If you're consistently 15% optimistic on collections, that bias is worth more than any single week's figure.

Look at the lowest point, not the closing balance. A quarter that ends at ₹8,00,000 but dips to −₹2,00,000 in week seven is a quarter you cannot survive. The trough is the number that matters.

Same discipline as running a proper spend-and-return sheet on your marketing. One sheet, updated on a schedule, that tells you the truth. If your ad spend is a meaningful line in your cash out, the marketing sheet that tells you whether it makes money belongs next to this one. Ad spend is paid weekly by card while the revenue it generates may collect in 90 days.

When borrowing helps and when it makes things worse

Credit is a legitimate tool for a timing problem and a terrible one for a margin problem. The test is simple. If your business is genuinely profitable and the shortfall is caused by the gap between paying and being paid, borrowing bridges a gap that will close. If you're unprofitable, borrowing buys time to lose more money. This is the business version of the test in good debt vs bad debt. What matters isn't the label on the loan but what the borrowed money does and whether the repayment survives a bad month.

For the profitable-but-squeezed case, a working capital limit such as an overdraft or cash credit facility priced against your receivables is the standard fit. You draw only what you need and pay interest only on the drawn amount. Invoice discounting converts specific approved invoices to cash immediately at a discount, and RBI-licensed TReDS platforms exist specifically so MSMEs can discount invoices raised on large buyers. A term loan repaid in fixed EMIs is the wrong shape for a fluctuating working capital need. The repayment doesn't flex when your collections do.

Compare the cost honestly. If discounting an invoice costs you 1.5% to get paid 75 days early, that's roughly 7.3% annualised. Cheap if the alternative is missing payroll. Expensive if the real fix was simply invoicing on the day of delivery instead of three weeks later. Fix the free things first, then price the paid things against what's left.

FAQs

Can a business be profitable and still fail?

Yes. It's one of the most common ways businesses die. Profit on paper doesn't pay salaries. Cash in the bank does. A business earning ₹5,00,000 of monthly profit that collects on 90-day terms can miss payroll while the P&L looks excellent. Manage cash as carefully as you manage profit.

What is more important, cash flow or profit?

Cash flow keeps you alive in the short term. Profit keeps you alive in the long term. Run out of cash and it's over, however profitable you looked. But a business with good cash flow and no profit, collecting deposits while losing money on every job, is just failing more slowly. You need both.

How much cash buffer should a small business keep?

Three to six months of fixed costs is the usual working range. Size it against your fixed outgoings, not revenue. Rent, salaries, EMIs and the tax you can't defer. If your cash conversion cycle is long or your customers are slow, sit at the top of that range or above it. If your revenue is concentrated in a handful of clients, that range is the wrong instrument entirely, and how much runway a services business actually needs works through the concentration-adjusted arithmetic instead.

Why is my profit high but my bank balance low?

Almost always working capital or non-P&L cash outflows. Check three things. How much customers owe you now versus six months ago. How much stock you're holding. And how much loan principal and tax you paid. Principal repayment and stock purchases drain cash without touching profit at all.

What is the difference between a P&L and a cash flow statement?

The P&L matches revenue and costs to the period the work happened in. The cash flow statement tracks money actually moving, split into operating, investing and financing activities. The same month can show ₹4,00,000 of profit and negative operating cash flow. That difference is the working capital swing.

Does taking a deposit reduce my profit?

No. A deposit changes only when cash arrives, not how much revenue you eventually book. Revenue is still recognised when you deliver. Until then the deposit sits as a liability. It's one of the very few levers that improves cash without touching price, cost or volume.

Key takeaways

Profit is an accounting figure matched to when work was done. Cash flow is money actually moving. Only cash pays salaries.
Profitable businesses go broke when the cash arrives after the bills. A timing problem, not a performance problem.
In the worked example, ₹30,00,000 of six-month profit produced a bank balance of −₹7,20,000, because ₹43,20,000 was absorbed by rising receivables net of payables.
Growth multiplies the working capital gap, so size the cash cost of a new contract using your cash conversion cycle before you sign it.
GST on invoices, TDS deductions and quarterly advance tax all take cash out ahead of collections, and none of them show up as a warning in your P&L.
Deposits, prompt invoicing, disciplined chasing and a rolling 13-week forecast fix most cash problems without a single new customer.

Billing Abroad, Spending at Home: FX Exposure for Indian Service Businesses

Prabhash Jha — Wed, 02 Sep 2026 09:25:05 +0000

You sign a twelve-month retainer with an overseas client at a fixed monthly fee in dollars. Everybody is happy. The rate feels good. Eight months later the same invoice, for the same work, at the same agreed number, lands in your account as noticeably less rupee than it did in month one — or noticeably more, which teaches you nothing and feels like skill.

Nobody repriced anything. You did not give a discount. Your client did not renegotiate. The contract was repriced by the currency market, monthly, in a direction neither party chose, and the only person carrying it was you.

Search this and page one is payment companies and FX platforms. Every one of them is technically correct and every one of them arrives at the same conclusion, which is that you should use their product. What is missing is the part that has nothing to do with a product: where you put the rate in the contract, how long you are actually exposed, and how to tell the difference between hedging a risk and simply charging for it.

What you are actually exposed to

Your exposure is not the gap between invoice date and payment date. It is the gap between the day you agreed the price and the last day you deliver at that price.

This is the single most common misunderstanding, and it is why FX advice written for goods exporters translates badly to service businesses. Three distinct exposures are in play and they are wildly different in size.

Transaction exposure. The window between raising an invoice and the money landing in your account. Real, measurable, and for most service businesses the smallest of the three. Every article you will read is about this one, because it is the one a payment product can address.

Pricing-period exposure. The window between agreeing a rate and stopping delivering at that rate. For a fixed-fee retainer this is the entire contract term plus however long the quote was outstanding before signature. It dwarfs transaction exposure and almost nothing sold to you addresses it, because the fix is a contract clause rather than a product.

Cost-base mismatch. Your revenue is in one currency and your salaries, rent and taxes are in another. This one never closes. It is not an event with a settlement date; it is a standing condition of the business, and it is the reason a currency move shows up in your payroll capacity rather than in a line on a bank statement.

A useful way to hold the distinction: transaction exposure is a risk you settle, pricing-period exposure is a risk you signed, and cost-base mismatch is a risk you are.

The number that matters: how long is your rate actually locked?

Add up the whole chain, because your real exposure duration is almost always several times longer than the payment terms everyone quotes.

The components:

Quote validity. How long between sending the proposal and signature. If your quotes have no expiry, this is unbounded, which is a genuinely bad place to start.
Contract term at the fixed fee. The full period before the price can change.
Payment terms. Net 30, net 45, whatever the invoice says.
Realisation lag. The gap between payment terms and actual receipt, which for overseas clients includes correspondent banking and, honestly, a client's own approval process.

Take a purely illustrative case to see the shape — these are made-up round numbers, not figures from any business: a quote outstanding for six weeks, a twelve-month term at a fixed monthly fee, net-45 terms, and two weeks of realisation lag. The last invoice of that contract is priced against a rate agreed roughly fourteen months earlier. Your exposure window is not 45 days. It is fourteen months, and it renews every time you re-sign without revisiting the rate.

Write your own version of that sum once. It reframes the problem from a banking question into a contracting question, which is where the answer actually lives.

Where to put the rate in the contract

Decide explicitly who carries the currency risk, and write it into the commercial terms rather than letting the invoice currency decide it by default.

Four structures, and there is no universally right answer — only a right answer for a given client, term and margin.

Structure	Who carries FX risk	Works when	The catch
Invoice in INR	The client	You have leverage, or the client has an India entity	Many overseas buyers will simply decline, and some cannot process it
Fixed fee in foreign currency	You, entirely	Short terms, or margin wide enough to absorb a move	The default, and the one that quietly reprices long retainers
Foreign currency with a rate band and reset clause	Shared	Long retainers with a professional counterparty	Needs a named reference rate and a defined trigger, or it is unenforceable in practice
Price in INR, invoice in foreign currency at an agreed reference rate on invoice date	The client, mostly	Clients who care about the total, not the currency	Your headline number moves month to month, which some buyers dislike

The third row is the one worth learning to sell, because it is the only structure that is genuinely fair over a long term and it is much easier to agree than people expect. The mechanics: name a specific published reference rate and source, define a band around the rate at signature, and state that if the rate closes outside that band for some defined period, either party may request a re-rate for future invoices. Not retroactively. Future invoices only.

Two details decide whether that clause is real or decorative. Name the rate source precisely — a specific published reference, not "the prevailing market rate", which resolves to an argument. And make the trigger symmetric. A clause that only protects you when the currency moves against you will be negotiated out, and rightly. A symmetric band gets signed, because it protects the client too, and because it makes you look like someone who has done this before.

The fourth row is underused and is often the easiest sell to a sophisticated buyer. Most clients are not attached to paying you in dollars. They are attached to knowing what the year costs. Pricing in rupees and converting at invoice date gives them a predictable annual budget in their own planning currency while removing the exposure from your side entirely.

Whichever you choose, this is a re-scope conversation with each client rather than a policy memo — the same mechanics as raising your prices without losing the accounts you want to keep, and it lands far better bundled into a renewal than sent as a standalone request.

Hedging versus pricing the risk in

These are two different strategies with different costs, and most small service businesses should be doing the second one.

Pricing it in means building an allowance for adverse movement into the rate you quote, and then simply accepting whatever the currency does. It costs you competitiveness on price. It costs you nothing in operational complexity, and it cannot go wrong in a way that requires a phone call.

Hedging means entering a contract — a forward, an option, a target-rate order — that fixes or bounds your conversion rate. It costs money, either explicitly as a premium or implicitly in the forward points. It also costs discipline, because a hedge against an exposure that then fails to materialise is a speculative position you did not intend to take. If a client cancels the retainer you hedged, the hedge does not cancel.

The decision rule I use, and it needs no numbers you have to invent: hedge when a plausible adverse move on a single contract would exceed the profit on that contract. Below that threshold you are buying protection against an outcome you can absorb, and paying for it in both cash and attention. Above it, one currency move can turn a delivered year of work into a loss, and that is exactly what hedging is for.

Two corollaries fall out of that rule and both are useful.

If the answer is "yes, hedge" on a majority of your contracts, the real finding is not that you need a treasury function. It is that your margins are too thin, and FX has merely been the thing that exposed it. That diagnosis is worth much more than the hedge.

And the cheapest hedge available to a service exporter is not a financial instrument at all. It is a natural hedge: deliberately holding foreign currency to pay foreign-currency costs. Most Indian service businesses have more of these than they realise — software subscriptions, cloud bills, ad platform spend, overseas contractors, conference travel. Every dollar of cost you pay from dollar revenue is a dollar you never converted twice and never paid a spread on. Which brings us to the account most service exporters should have and many do not.

The EEFC account, and its one-month limit

An Exchange Earners' Foreign Currency account lets you hold export earnings in foreign currency in India instead of converting on arrival, and it is the most useful FX tool available to a small service exporter.

Per the RBI's own FAQ on EEFC accounts, the essentials are:

100% of foreign exchange earnings can be credited to the account.
It is a current account only, and no interest is payable on the balance.
Cheque facility is available for operating it.
Permitted debits include a range of current and capital account transactions, payments to Indian residents for goods and services, and customs duties.

What it buys you in practice is twofold. It lets you pay foreign-currency costs out of foreign-currency revenue, which is the natural hedge above. And it gives you some control over when you convert, rather than converting at whatever rate happened to prevail on the day a client's payment cleared.

Now the constraint, which is the part that gets left out of the enthusiastic write-ups. The RBI's rule is that the sum total of accruals in the account during a calendar month must be converted into rupees on or before the last day of the succeeding calendar month, after adjusting for utilisation on approved purposes. So the timing flexibility is real and it is bounded. It is a window measured in weeks, not a foreign-currency reserve you can sit on through a bad quarter.

That distinction matters because a strategy built on holding through an adverse move will hit the conversion deadline and convert anyway, at the worst possible moment, having felt like a plan the whole way. Use the account for the natural hedge and for modest timing discretion inside the permitted window. Do not use it as a substitute for pricing the risk in.

The compliance layer that changes your cash timing

Two rules can turn an FX question into a tax and cash-flow question, and both have deadlines attached.

GST treats export of services as zero-rated, but only if you are paid in convertible foreign exchange. Under Section 2(6) of the IGST Act, a supply qualifies as an export of services only when a set of conditions is met, and receipt of payment in convertible foreign exchange — or in rupees where the RBI permits it, such as through a Special INR Vostro arrangement — is one of them. To export without paying IGST upfront you furnish a Letter of Undertaking on Form GST RFD-11 under Rule 96A, and that undertaking includes receiving payment within one year of the invoice date. Which means a slow-paying overseas client is not only a cash-flow problem. It is a compliance clock.

FEMA sets a period within which export proceeds must be realised and repatriated, and this is one to check rather than remember. The realisation period has moved more than once recently, in both directions — it was extended, then brought back in. Any blog post asserting a specific figure, including one written last quarter, may be describing a rule that has since changed. The operational instruction is: read the current RBI Master Direction on Export of Goods and Services, or ask the banker who handles your inward remittances, before relying on a number. That is not a hedge on my part; it is the correct professional habit for any rule that has changed twice in two years.

Neither of these is really about currency. They are about the same thing that catches every services business — an obligation whose timing is set by somebody else's payment behaviour. That is the identical mechanism behind GST and TDS hitting your cash before the client does, and if you bill overseas clients you now have both patterns operating at once.

The reconciliation nobody does

Reconcile each inward remittance against the contract, because the fee you are paying for the conversion is inside the rate rather than on the statement.

This is the operational habit that changed the most for me, and it takes about ten minutes per remittance. Four numbers per payment:

The contracted amount in the invoice currency.
The amount actually remitted in that currency, which can differ — some clients deduct their own bank's charges, and some deduct withholding.
The rupees credited to your account.
A published reference rate for that currency pair on the credit date.

Then divide the rupees credited by the foreign currency received. That is your realised rate. Compare it to the reference rate. The difference is the all-in cost of the conversion, and it is almost always larger than any explicitly disclosed charge, because most of the margin lives in the spread rather than in a fee line.

Do this for three months and you will know two things you almost certainly do not know today: what your effective conversion cost actually is, and whether it varies by amount, by day, or by who at the bank happened to process it. That is a negotiable number once you can state it. It is invisible until you calculate it, which is exactly why nobody negotiates it.

Keep the record of every remittance and its documentation together, too. You will need it for GST substantiation and for FEMA realisation evidence, and assembling it retrospectively at year end is one of those tasks that costs five times what it should. The general principle — that the reconstruction is only possible if you kept the inputs — is the same one behind cash flow versus profit: the P&L can look fine while the bank account tells a different story, and only the line-by-line record reconciles them.

The practitioner sequence

If you bill overseas clients and have never treated this as a system, do these in order.

Calculate your real exposure duration on your largest contract. Quote validity plus term plus payment terms plus realisation lag. This one sum reframes everything else.
Put an expiry date on every quote. The cheapest fix on this list and the one most often missing.
Open an EEFC account if you have any foreign-currency costs at all, and start paying them from it.
Reconcile three months of remittances using the four numbers above. Find out what conversion actually costs you.
Take that number to your bank, and to one alternative provider. It is a rate, and rates are negotiable when you can quote them.
Add a symmetric rate-band clause to your next long retainer renewal. Not to the current ones mid-term.
Apply the hedging rule — would a plausible adverse move exceed the profit on this contract? — to each contract individually, and hedge only where the answer is yes.
Check the current realisation period and your LUT status for the year. Both are deadline-bearing and neither announces itself.

Steps 1 through 4 cost nothing but attention and produce most of the benefit. Steps 5 through 8 are where the actual money is, and none of them are available until you have done the first four.

When FX is not your real problem

If a plausible currency move can turn a contract unprofitable, you do not have a currency problem. You have a margin problem, and the currency found it.

This is worth sitting with, because it is the most valuable thing FX analysis produces and it is never the thing people set out to learn. Currency movement is a revealer. It applies a small, random pressure to every contract you hold, and the ones that break under it were already close to the line for reasons that have nothing to do with the exchange rate — mispriced at the outset, or eroded since by scope that grew while the fee stayed still.

The erosion mechanism is worth understanding on its own terms; how a profitable retainer quietly becomes an unprofitable one works through it as arithmetic rather than grievance. And if the honest answer is that the original price was too low, that is a pricing conversation, not a treasury one — how to price your services as a freelancer or consultant is the right place to start.

A business with real margin experiences currency movement as noise in the monthly numbers. A business without it experiences the same movement as an existential event. The difference is not the hedging strategy.

FAQs

Should I invoice international clients in INR or in their currency?

Whichever currency you invoice in, someone is carrying the exchange risk, and the invoice currency decides who. Invoicing in rupees moves it to the client, which many overseas buyers will decline or cannot operationally process. The middle path that gets signed most often is to price in rupees and invoice in the client's currency at a named reference rate on invoice date, which gives them a predictable budget while keeping the exposure off your books.

What is an EEFC account and should a service exporter have one?

It is a non-interest-bearing current account, held in India in a freely convertible foreign currency, that lets you retain export earnings without converting them immediately. The RBI permits 100% of foreign exchange earnings to be credited to it. It is worth having if you have any foreign-currency costs at all, because paying those costs from foreign-currency revenue avoids converting twice and paying a spread each time.

How long can I hold foreign currency in an EEFC account?

Not indefinitely. The RBI requires that the total accruals in a calendar month be converted into rupees on or before the last day of the following month, after adjusting for permitted utilisation. So the account gives you genuine timing discretion measured in weeks, not a reserve you can hold through an adverse quarter. Any strategy that depends on holding longer will hit the deadline and convert anyway.

Do I need to hedge currency risk as a small service business?

Usually no, and the useful test is whether a plausible adverse move on a single contract would exceed the profit on that contract. Below that, you are paying cash and attention to insure against something you can absorb. Above it, hedge. If the answer comes out "yes" across most of your contracts, the finding is that your margins are too thin rather than that you need a treasury function.

Is export of services zero-rated under GST?

Yes, subject to conditions set out in Section 2(6) of the IGST Act, one of which is that payment is received in convertible foreign exchange, or in rupees where the RBI specifically permits it. To export without paying IGST upfront you file a Letter of Undertaking on Form GST RFD-11 under Rule 96A, which carries an undertaking to receive payment within one year of the invoice date. A slow-paying overseas client therefore creates a compliance exposure as well as a cash one.

How long do I have to receive payment from a foreign client under FEMA?

There is a defined realisation and repatriation period, and you should check the current RBI Master Direction on Export of Goods and Services rather than rely on a figure from an article. The period has been changed more than once in recent years, in both directions, so material written even a few months ago can be describing a superseded rule. Your bank's trade desk will confirm the current position.

How do I find out what my bank is actually charging me to convert currency?

Divide the rupees credited by the foreign currency actually received, then compare that realised rate against a published reference rate for the same date. The gap is your all-in conversion cost, and it is normally much larger than any disclosed fee because most of the margin sits in the spread. Do this for three months and you will have a number you can negotiate with. Until you calculate it, it is invisible.

What should go into a currency clause in a long-term contract?

A named published reference rate and its source, a defined band around the rate at signature, a trigger stating how far and for how long the rate must move outside that band, and a statement that any re-rate applies to future invoices only. Make the trigger symmetric so it protects both sides — a one-sided clause gets negotiated out, and a symmetric one signals that you have run this before.

Key takeaways

Your exposure is the price-agreed-to-delivery window, not the invoice-to-payment window. It is usually months longer.
Add it up once: quote validity plus term plus payment terms plus realisation lag.
The invoice currency decides who carries the risk. Choose it deliberately.
Price in INR, invoice in the client's currency at a named reference rate. It is the structure most likely to be signed.
A rate-band clause must name the rate source and be symmetric, or it will not survive negotiation.
Hedge only when a plausible adverse move would exceed the profit on that contract.
The cheapest hedge is a natural one: pay foreign-currency costs from foreign-currency revenue.
An EEFC account holds 100% of export earnings, pays no interest, and must be converted by the end of the following month. Weeks of flexibility, not a reserve.
Zero-rated GST on exported services requires payment in convertible foreign exchange, and Rule 96A puts a one-year clock on it.
Check the current FEMA realisation period rather than trusting any published figure. It has changed twice recently.
Compute your realised rate against a reference rate. The conversion cost is in the spread, not on the statement.
If a currency move can make a contract unprofitable, the contract had no margin. FX only found it.

Attribution Windows Are a Commercial Term, Not a Technical Setting

Prabhash Jha — Wed, 02 Sep 2026 09:25:04 +0000

A partner asks you to move the click window from 7 days to 30. It arrives as a small technical request, usually late in an email, usually phrased as a fix. Our conversions aren't being credited properly. Can you extend the window?

Nothing about that sentence is technical. What has just been requested is a price increase, and the increase is invisible because it doesn't appear in the rate. The commission stays at whatever it was. The volume of sales stays exactly what it was. What changes is how many of those sales get labelled as somebody's work, and every sale that changes label moves money.

I have agreed to window extensions I shouldn't have, because the request was framed as a measurement correction and I answered it as one. The pattern is consistent enough to be a rule now: treat the window as a term in the contract, priced and traded like any other. Not as a dropdown.

What an attribution window actually decides

An attribution window does not decide what happened. It decides who gets paid for what happened.

This distinction sounds pedantic and it is the whole post. A customer clicked a partner's link on Tuesday, saw two of your ads on Wednesday, searched your brand name on Friday, and bought on Saturday. That sequence is a fact. It doesn't change when you change a setting. What changes is which of those four touches your system holds responsible on payout day.

So when someone tells you a longer window is "more accurate", ask accurate about what. A 30-day window is more generous to the earliest touch. A 1-day window is more generous to the last. Neither is measuring the customer's intention, because nobody can. Both are allocation rules, and an allocation rule is a commercial policy wearing a technical costume.

The second thing the window decides is subtler and it costs more. In every channel where an algorithm optimises against your conversion feed, the window defines what the algorithm is chasing. Google's documentation is direct about it: Smart Bidding will "count and optimize conversions for any window you choose". Widen the window and you have not just changed a report. You have changed what the bidding system believes a good click looks like, and it will go and buy more of that.

Who gains when the window gets longer, and who pays for it

A longer window transfers credit from partners who close to partners who introduce, and the bill lands on whoever holds the margin.

That transfer is not evenly distributed. It reliably favours some partner types and reliably harms others, and knowing which is which is most of the negotiation.

Partner type	Effect of a longer click window	Why
Content, review, comparison sites	Gains substantially	Their touch happens during research, days before intent matures
Newsletter and community placements	Gains	Read now, buy at the weekend
Coupon and voucher sites	Gains modestly, and it is the least deserved gain	Their touch is usually at checkout anyway; a longer window mostly protects them from being deduped out
Loyalty and cashback	Gains	Same checkout-adjacent position, now insulated
Retargeting and display	Gains heavily	They serve to people already in your funnel, so a wide window catches almost everything
Brand-term search bidders	Gains heavily, and this is the one to watch	Bidding on your own name intercepts demand you already created
Influencer and social, one-off posts	Gains	Discovery-shaped, long lag to purchase
Your own email and organic	Loses	These get deduped out by paid touches with wider windows

Read the last row again, because it is the one nobody puts on a slide. The party that pays for a wider window is frequently not another partner. It is your own unattributed and owned demand — the traffic you were going to get anyway, which now arrives with an invoice attached.

That is the whole reason a window extension can raise your blended cost of sale without raising a single rate, and without a single additional order. If you have never sat down and worked out what an order actually has to leave behind, the sheet that tells you whether your marketing makes money is the arithmetic this decision sits on top of.

Three numbers that tell you what a longer window costs, before you agree to it

You can price a window extension before agreeing to it, using data you already hold. Three numbers, in this order.

1. Your time-to-conversion distribution. Not the average. The distribution. What share of conversions land within 1 day, 7 days, 14, 30. Every major platform will show you this, and most affiliate networks will too if you ask the account manager for a lag report. If 85% of conversions already land inside 48 hours, a 30-day window is buying you almost nothing in truth and a lot in liability, because the extra 28 days are mostly catching coincidence. If the distribution has a genuine long tail — considered purchases, high ticket, B2B — the partner asking has a real case.

2. The overlap rate. Of the conversions that a longer window would newly credit to this partner, how many are already being credited to someone else, including to your own channels? This is the number that turns the conversation from generous to specific. A longer window rarely finds new sales. It mostly relabels existing ones. Both Google Ads and Meta let you model this without touching your live configuration — Meta's Compare Attribution Settings sits in the Columns dropdown in Ads Manager and shows how conversions redistribute across windows on data you have already collected. Use it before the call, not after.

3. The new-customer share. Of the sales this partner is credited with, what fraction are first-time buyers? A partner whose credited volume is mostly repeat purchasers is being paid to stand near your existing customers. A longer window makes that worse, because repeat buyers were already coming back. This single ratio has told me more about a partner's real contribution than any conversion count, and it is the number partners are least keen to discuss.

None of the three requires new tooling. All three require that your tracking is actually intact, which is a much less safe assumption than it sounds — affiliate tracking breaks quietly, and payout day is the wrong time to find out covers the failure modes that make these numbers lie before you ever get to interpret them.

Why platform defaults are not neutral

Every platform default encodes a commercial position, usually the platform's own.

Look at what the defaults actually are. In Google Ads, per Google's conversion window documentation, a new conversion action defaults to a 30-day click-through window, configurable from 1 up to 30, 60 or 90 days depending on the conversion source. The engaged-view window defaults to 3 days and the view-through window to 1 day, both adjustable to 30. On Meta, the default for campaigns optimising to website conversions is a 7-day click alongside shorter engage-through and view-through windows.

Then look at the other end of the spectrum. Amazon Associates runs one of the shortest windows in the industry — a 24-hour cookie, extended only if the customer adds the item to their cart within that day, in which case the purchase can still credit up to 90 days later, and only for the items placed in the cart during that original session.

Those are three completely different commercial philosophies presented as configuration. Amazon's window is short because Amazon has demand of its own and does not need to pay you for it. A platform selling you media has every reason to default wide. Neither is lying. Both are negotiating, and one of them has already won because you never opened the setting.

There is one more thing in Google's documentation worth holding on to, because it changes the timing of any negotiation: window changes apply only going forward. Shorten a 30-day window to 10 and the 10-day rule applies to conversions recorded from that day on. Past conversions are not restated. Which means you cannot quietly fix a bad window retroactively, and it means a partner who wants the change made before a big sale period is asking for something with a specific value they have probably already calculated.

The window is one term in a set of four. Never move it alone.

Never concede a window extension on its own, because the window is one of four interlocking terms and moving one without the others is how you pay twice.

The four:

Window length — how long after the touch a conversion still credits.
The validation or reversal window — how long you retain the right to claw back a credited sale for a return, a cancellation, a chargeback or a fraud finding.
The deduplication rule — what happens when two partners both have a live claim on the same order, and which position wins.
The rate, including tiers and any new-customer differential.

These trade against each other cleanly, and the trades are where the actual negotiation lives.

Give a partner 30 days instead of 7, and take a longer validation window in exchange — a wide credit window with a short reversal window is the single worst combination to sign, because it maximises the volume you pay on and minimises your right to correct it. Give 30 days, and take a new-customer differential: full rate on first-time buyers, reduced rate on repeats. Give 30 days to genuine content partners, and simultaneously tighten the dedup rule so that a coupon site landing at checkout cannot outrank an introducer who did the work on day one.

That last one deserves saying plainly. In most default configurations the last touch wins, which means a longer window given to everyone hands the largest gain to the partner closest to the checkout button. If you extend the window as a blanket policy, you are usually paying more to the partners who contributed least. The extension has to be paired with a dedup change or it inverts your own intention.

How to write it into the contract

Put the window in the commercial terms next to the rate, not in a technical appendix, and name all four terms in the same clause.

What that looks like in practice — described rather than drafted, because your counsel should write the words:

State the window in days and name the trigger. Days from a qualifying click, defined. Not "standard industry window", which resolves to whatever the network's default happens to be on the day of the dispute.
State the validation period and what may be reversed within it. Returns, cancellations, chargebacks, duplicate orders, orders failing a fraud check, orders placed by the partner's own staff. Silence here defaults to the network's terms, which were not written for you.
State the deduplication rule explicitly, including how your owned channels are treated. If your email and organic traffic are eligible to win a contested order, say so. Most contracts are silent, and silence is a decision that goes against you.
State what happens to in-flight touches when the window changes. Because changes are forward-only in most systems, a mid-month change creates a genuinely ambiguous cohort. Decide in advance who owns it.
Give both sides a review right at a fixed cadence. A quarterly review clause turns "we need to talk about your window" from an accusation into a diary entry. This is the same principle as putting a quarterly access review in an agency contract, and it works for the same reason.
Tie the window to a performance floor, not to tenure. A partner keeps the wider window while reversal rate and new-customer share stay inside agreed bands. This is the clause that lets you be generous without being permanently exposed.

If you are on the other side of this table — choosing programmes to promote rather than running one — the same four terms are what you should be reading before you join anything. The checklist for choosing an affiliate programme covers the partner-side view, and it is worth reading both directions before a negotiation.

The conversation, in order

Run the renegotiation in a fixed sequence, because the order determines whether you are discussing evidence or trading opinions.

Ask for the lag data before you respond at all. "Send me your time-to-conversion distribution for the last full quarter." A partner with a real case has this and will send it. A partner testing you will send an anecdote about a customer who bought three weeks later.
Bring your own overlap number. Model the redistribution first. Walking in with "extending this window moves credit on roughly this proportion of orders, most of which currently credit elsewhere" ends the accuracy framing permanently and moves the conversation to price.
Name the trade before they name a number. "Thirty days works, paired with a longer validation window and a new-customer rate." Concede the thing they asked for, in the same sentence as the thing you need.
Agree a review date and the bands. What reversal rate, what new-customer share, reviewed when.
Change one term, then wait a full purchase cycle. Two simultaneous changes make the result uninterpretable. This is the same discipline as never changing a tag and a tracking configuration in the same week.
Write down what the numbers looked like the day before the change. Nobody does this and everybody needs it. See the next section for why.

What to do when a partner insists on a window you cannot afford

When the window they need is genuinely wider than your margin supports, change the shape of the deal rather than the window.

Options, roughly in order of how often they work:

Tier the commission by time-to-conversion. Full rate inside 7 days, reduced rate from 8 to 30. This is the honest expression of what a longer window is worth: something, but less. It is also the offer that separates partners who want fair credit from partners who want a raise.
Pay a flat placement fee plus a short window. For discovery-shaped partners — content, newsletters, one-off influencer posts — the fee compensates the introduction directly, and the short window stops it being paid for twice.
Differentiate on new customers only. Wide window on first-time buyers, narrow on repeats. Cleanly aligned with what you are actually buying.
Exclude branded search from the wider window. If a partner bids on your own brand terms, a wide window is close to pure leakage. Carve it out explicitly.
Offer the wider window as a time-boxed pilot with an agreed measurement plan. One quarter, defined metrics, an explicit end date rather than an automatic renewal. Most window extensions that turned out badly for me became permanent by default rather than by decision.

And sometimes the answer is no, and the partner leaves. That is a real outcome and it is survivable. A partner whose volume is only profitable under an allocation rule you would not otherwise choose was not profitable. It just looked that way.

The measurement trap: a wider window makes everything look better at once

The day after you widen a window, every partner's conversion count goes up, your reported blended ROAS improves, and none of it means anything.

This is the trap that catches people who did everything else right. You made a careful, negotiated change. Then you looked at the numbers the following month and they were better, so you concluded the change was good. But a wider window mechanically increases attributed conversions across the board, because more historical touches are now eligible. Nothing about the business changed. The counting changed.

Three defences, all cheap:

Snapshot the before. Export the full set of per-partner metrics — conversions, reversal rate, new-customer share, blended cost of sale — on the day before the change. Not a summary. The export. This costs ten minutes and it is the only version of the baseline that will still exist in three months.

Judge the change on unattributed totals. Total orders and total revenue don't care about your allocation rule. If total orders are flat and attributed conversions rose 20%, you have bought relabelling, not growth. That comparison is the entire test, and it takes one line.

Never compare across a window change. A pre-change month and a post-change month are different units of measurement wearing the same name. If you must report a trend across that boundary, say in the report that the boundary exists. The habit of stating which measurement regime a number came from is the same one that stops Search Console numbers quietly misleading you — in both cases the number is correct and the comparison is not.

The broader version of this discipline — deciding what a metric is for before you optimise against it — is the spine of the performance marketing playbook I actually use. Attribution windows are just the place where the cost of getting it wrong is easiest to measure and hardest to notice.

FAQs

Is a longer attribution window more accurate?

No. It is more generous to earlier touches. Accuracy would require knowing what actually caused a purchase, which no window can tell you. A window is an allocation rule, and the honest question is not which one is true but which one you want to pay against. Pick the one that matches your real time-to-conversion distribution, then treat it as a commercial position rather than a measurement.

What is a reasonable attribution window for an affiliate programme?

The one that covers the bulk of your actual conversion lag, which you should measure rather than benchmark. Industry norms range from Amazon's 24 hours to 90 days on many SaaS programmes, and that spread exists because the underlying purchase behaviour genuinely differs. If most of your conversions land inside a couple of days, a long window is buying liability rather than fairness.

Does changing the conversion window affect historical data?

Not in Google Ads. Google's documentation states that window changes apply only going forward — shorten a 30-day window to 10 days and the shorter rule applies to conversions recorded from that day onward, with past conversions left as recorded. Practically, this means you cannot retroactively repair a window you regret, and it means the timing of any change is itself negotiable value.

Who actually loses when an affiliate window is extended?

Usually your own owned channels, and any partner positioned earlier in the journey than the one whose window widened. Email, organic and direct traffic get deduped out of contested orders by paid touches with wider eligibility. Because those channels do not send invoices, the loss shows up as a rising blended cost of sale with no obvious cause.

Should the reversal window match the attribution window?

They serve different purposes, so they need not match — but the reversal window should never be materially shorter than the attribution window. A wide credit window paired with a short reversal window means you pay on the maximum possible volume while holding the minimum possible right to correct it. That combination is the most expensive thing you can sign, and it is common because the two terms are usually negotiated by different people at different times.

How do I model the impact before making the change?

Use the comparison tooling already in the platforms. Meta's Compare Attribution Settings, in the Columns dropdown in Ads Manager, redistributes conversions across windows on data already collected without altering your live configuration. Google Ads exposes days-to-conversion reporting for the same purpose. Between them you can price the change on historical data, which is a much stronger position than agreeing first and measuring after.

A partner says their conversions are not being credited properly. Is that a tracking problem or a window problem?

Rule out tracking first, because the two look identical from the partner's dashboard and only one of them costs you money to fix. Missing sub-IDs, a broken postback, a redirect stripping parameters and consent-blocked tags all present as under-credited conversions. Ask for the lag distribution: if their conversions are landing inside your existing window and still not crediting, it is a tracking fault, and extending the window would have hidden a bug behind a payment.

Key takeaways

An attribution window decides who gets paid, not what happened. It is a commercial term.
A wider window rarely finds new sales. It relabels existing ones, often away from your owned channels.
Price it with three numbers first: time-to-conversion distribution, overlap rate, new-customer share.
Platform defaults are commercial positions. Amazon's 24 hours and a 30-day media default are both arguments.
Never move the window alone. Trade it against the validation window, the dedup rule and the rate.
A wide credit window with a short reversal window is the worst pair to sign.
Extending the window for everyone pays the checkout-adjacent partners most. Pair it with a dedup change.
Window changes are forward-only, so you cannot fix a regret retroactively. Timing is value.
Snapshot every per-partner metric the day before you change anything.
Judge the result on total orders, not attributed conversions. Attributed conversions always rise.