DEV Community: Henil Mehta

What Would You Choose: ChatGPT, Grok, Copilot or a Custom AI Agent?

Henil Mehta — Tue, 13 Oct 2026 19:22:08 +0000

You already pay for ChatGPT and Grok. Why would you need a custom AI agent?

You already pay for ChatGPT and Grok. Why build an agent?

Picture the owner of a packaging unit in Rajkot. He pays for ChatGPT, his son uses Grok, and the accounts team got Copilot with Microsoft 365. Between them, they draft emails, summarise PDFs and answer half his questions before lunch, and he's right that they help the business.

So when someone pitches him a custom AI agent, his reaction is fair: "I already pay for these. Why would I need one more?"

On the custom AI agent vs ChatGPT question, I mostly agree with him, even though I build custom agents. These tools got a lot more capable this year. What matters is where they stop, and whether your work lives past that line.

What ChatGPT, Grok and Copilot are genuinely good at now

Credit first. Most pages comparing a custom agent with ChatGPT still describe ChatGPT as a chatbot that answers questions. That stopped being true this year.

OpenAI Dots. OpenAI launched Dots at DevDay on 29 September 2026. These are always-on agents inside ChatGPT, each with its own cloud computer and browser. They run in the background, learn your preferences, run on GPT-6 Astra, and you can reach them from ChatGPT, Slack or Teams. OpenAI says Dots "can readily connect to over 4,000 apps" through plugins, and it runs an automatic review before actions that touch your accounts or share information. OpenAI also says, fairly, that it "can still make mistakes". It's rolling out to Pro and Business Premium, with Enterprise in beta.

xAI Grok Bot. Grok Bot launched in beta on 11 August 2026. Its agents sign in to your company software and do multi-step tasks on a cloud virtual machine. There are 220 plugins, including Google Workspace, Slack, Microsoft 365, Salesforce, Notion, GitHub and Jira, and none are installed by default, which I like. It pauses before sending, publishing, buying or deleting, though an "Always Allow" setting skips those pauses. Worth knowing: each account gets one VM, shared by all its bots, logins included. To be fair about scale, Grok has the fewest paid subscribers of the three, and Grok Bot is a two-month-old beta, so far open to some SuperGrok Heavy and Cursor subscribers, with teams promised access later.

Microsoft Copilot. Copilot (formerly Microsoft 365 Copilot) passed 30 million paid seats by Microsoft's FY26 Q4 results on 29 July 2026. Microsoft says your prompts and Graph data aren't used to train its foundation models, and Copilot is covered by its Data Protection Addendum. Copilot Studio goes further: you can build custom connectors to "any publicly available API". If your company already lives in Outlook, Excel and Teams, that's a strong place to start.

For a rough sense of who uses what: Recon Analytics estimated that in January 2026, ChatGPT had 55.2% of paid AI subscribers, Copilot 11.5% and Grok 3.4% (as reported by Stackmatix). It's one firm's estimate, so read it as a direction.

If your work is drafting, research, summarising and acting inside mainstream cloud apps, one of these may be all you need. I'd try that first.

Where they stop: six honest limits

None of these is a bug. A product built for millions of companies can't know how yours runs.

Your systems behind the firewall. The 4,000 apps Dots connects to are popular cloud apps. Tally on the office PC, a 12-year-old ERP on a local server or your internal database usually aren't on that list. Boomi's ChatGPT setup guide puts it plainly: "ChatGPT connects to remote MCP servers over the public internet. Servers behind VPNs or private networks will not connect." You can still reach those systems, with a local connector or a safely exposed endpoint. Copilot Studio also needs a data gateway plus build work to reach on-premises systems. Either way it's custom engineering, on someone's platform.
Your rules, beyond generic approvals. Dots reviews actions before they touch your accounts, and Grok Bot pauses before it sends or deletes. That's a sensible default. What it doesn't know is that your sales team can give, say, 5% off on their own and anything more needs your sign-off, or that no quote leaves without a GST number and a delivery date. In a custom agent I put those rules in plain code that runs every time, so the model never gets to judge whether it followed them.
Token bleed. Agents read, plan, call a tool, read again, and every step costs tokens. In one 2026 study of coding agents (Bai et al., summarised by Stanford's Digital Economy Lab), agentic tasks used about 1,000 times more tokens than code chat, and the cost of the same task varied by up to 30 times. That was coding, so the exact figures won't carry over to your business, but an always-on agent doing everything with one large model gets expensive at volume. A custom build can send narrow steps, like reading an order number or sorting an enquiry, to a smaller, cheaper model and save the big one for the hard parts.
Privacy and India's DPDP rules. The business tiers of all three don't train on your data by default. The quieter risk is staff pasting client data into personal Free, Plus or Pro accounts, where training is on by default unless they opt out. India's DPDP Rules were notified in November 2025, and the core duties (notice and consent, security safeguards, breach notification, cross-border transfers) apply from 13 May 2027. Also, OpenAI's India data residency covers data at rest, and inference ran only in the US as of November 2025. I'm not a lawyer and this isn't legal advice, but if client data is involved you should know where it's stored and where it's processed. A custom build lets you choose both.
Lock-in. Dots runs on OpenAI's models and Grok Bot runs on Grok. If a better or cheaper model comes out next quarter, you wait for your vendor. Copilot is fairer here, since it already works with more than one model family. A custom agent lets you pick a model for each task and swap one out later without rebuilding the rest.
Your edge. This one is opinion. Your competitor can subscribe to the same bot tomorrow, with the same features. What they can't buy is the way you price, qualify and follow up, built into an agent that runs on your data. If a workflow is part of why customers pick you, I'd rather own it than rent it.

Custom AI agent vs ChatGPT, Grok Bot and Copilot, side by side

Here's the comparison row by row. By off-the-shelf I mean Dots, Grok Bot and Copilot.

Best for. Off-the-shelf: personal productivity and tasks inside popular cloud apps. Custom: one repeated business workflow that runs across your own systems.
Connects to. Off-the-shelf: thousands of popular apps through plugins, and Copilot Studio adds connectors to public APIs. Custom: those, plus on-premises and legacy systems through connectors built for them.
Approval rules. Off-the-shelf: built-in pauses before risky actions (Grok Bot lets you switch them off with Always Allow). Custom: your business rules checked in code, plus a person's approval wherever you want one.
Model choice. Off-the-shelf: the vendor's models, with Copilot offering more than one. Custom: a model per step, replaceable later.
Data location. Off-the-shelf: set by the vendor's plans and regions. Custom: you choose the provider and region, and can keep sensitive steps on your own servers.
Cost control. Off-the-shelf: a subscription, with usage handled the vendor's way. Custom: you pay for what each step uses and can route simple steps to cheaper models, but you also pay to build and maintain it.

When is ChatGPT enough, and when is custom worth it?

ChatGPT, Grok or Copilot is probably enough when:

The work is drafting, research or summarising for you and your team.
Your tools are mainstream cloud apps: Gmail, Google Sheets, Slack, Microsoft 365, a cloud CRM.
Volume is low and a mistake is cheap to fix.
A person reads the output before it goes anywhere.

A custom agent is worth it when:

The workflow needs on-premises or legacy systems: Tally, a local ERP, an internal database.
It runs hundreds of times a week, so token costs add up.
Money or customer data is at stake.
You have rules a generic approval step can't know about.
It spans several systems in several steps.

Tip: If you can't write the workflow down on one page, don't build an agent for it yet. Fix the process first. And if it happens five times a month, ChatGPT or Copilot with a person checking the result is fine.

Example: from IndiaMART enquiry to sent quote, step by step

Here's the kind of workflow where I'd go custom. Picture a Rajkot supplier of brass fittings. This is an illustration, not a client.

An enquiry arrives on IndiaMART. The agent picks it up and reads the product, quantity and delivery city.
It checks stock and price in the ERP or Tally on the office server through a local connector, so that system stays off the public internet.
It drafts a quote in your format, with your terms.
Code checks the draft against your rules: discount within the limit, quantity above the minimum order, required fields filled. If anything fails, the quote goes to a person instead.
The owner gets the quote on WhatsApp and approves or edits it.
Once approved, the agent sends the quote to the buyer and logs it with the enquiry in the CRM.

I'm not claiming Dots or Grok Bot can or can't do this with your Tally setup. The local connector and the rule checks are what make it yours, and they're custom work on any platform. We've written more about the IndiaMART side at https://dizilus.com/indiamart-lead-automation/.

If custom, who builds it?

You have three realistic routes: your own developer, an agency, or Copilot Studio if you're deep in Microsoft and your systems can be reached. I've compared the agency and in-house routes honestly at https://dizilus.com/blog/automation-agency-vs-in-house/.

Whoever builds it, ask three questions. Which model runs each step? What happens when the model gets something wrong? Who fixes it when it breaks on a Saturday?

Start with one workflow

Don't start by picking a tool. Pick one repetitive workflow, write down every step, and mark which systems it touches and which steps need a person's judgement.

If it all lives in mainstream cloud apps, try Dots, Grok Bot or Copilot first. If it touches your own systems or your own rules, that's the kind of agent we build. See how we work at https://dizilus.com/ai-agents/, or email the workflow to tea@dizilus.com and I'll tell you honestly which route fits.

FAQ

Can ChatGPT, Grok or Copilot connect to my old ERP or Tally?
Usually it takes extra work. Their plugins cover popular cloud apps, and ChatGPT connects to remote MCP servers over the public internet, so a system behind a VPN or on an office server needs a connector or gateway built for it. Copilot Studio can reach on-premises systems with a data gateway and some build work. Either way, that's custom work.

Is my business data used to train their models?
On the business tiers of ChatGPT, Grok and Copilot, not by default. On personal plans like ChatGPT Free, Plus or Pro, training is on by default unless you opt out, so keep client data out of personal accounts.

Why do AI agents cost more to run than chat?
An agent plans, calls tools and rereads results, and every step uses tokens. One 2026 study of coding agents found agentic tasks used about 1,000 times more tokens than code chat, with the cost of the same task varying by up to 30 times. A custom build can send simple steps to smaller, cheaper models.

When is ChatGPT enough for my business?
When the work is drafting, research or tasks inside mainstream cloud apps, volume is low, and a person checks the output before it goes anywhere. Plenty of businesses never need more than that.

Can I switch AI models later?
With Dots or Grok Bot you use the vendor's models. Copilot already works with more than one. A custom agent can use a different model for each step, and you can swap one out later without rebuilding the rest.

What does a custom AI agent do that a ChatGPT agent can't?
It connects to your own systems, including on-premises ones, checks your business rules in code before it acts, lets you choose a model for each step, and keeps data where you decide. Dots and Grok Bot are improving fast, so judge them on your actual workflow.

Originally published at dizilus.com.

How we wire up lead replies that go out in under a minute

Henil Mehta — Sat, 10 Oct 2026 09:55:29 +0000

A fast reply isn't just "thanks, we'll be in touch". Within a minute of an enquiry arriving, a useful reply should:

answer what the person asked, or at least show you read it
ask one qualifying question, or offer a time to talk
log the lead and the reply in your CRM
tell the right person when a human is needed

Here's the stack we use to get there.

1. Your CRM's own automation

If you already have a CRM, start here. The lead data is already in it.

HubSpot workflows (Professional and Enterprise tiers of hubs like Sales and Marketing): a form submission or new contact can send an email, create a task, assign an owner and notify the team.
Zoho CRM workflow rules run when a record is created or edited. Instant actions include emails, tasks, field updates, webhooks and custom functions.
Pipedrive automations (Settings → Automations): a new deal or person can trigger an email from your synced mailbox, an activity or a field update.

The limit is the reply itself. Most CRM automations send a template with the name merged in. Fast, but it can't respond to what the person wrote. Also check which plan the automation features need before designing anything; they often sit on higher tiers.

2. WhatsApp, if leads message you there

The WhatsApp Business app on a phone is fine for a few chats, but it isn't built for automation. For automatic replies you need the WhatsApp Business Platform (the "API"), usually through a provider:

Interakt: popular in India; shared inbox, no-code chatbot builder, lead management, Shopify and WooCommerce integrations.
WATI: shared team inbox, no-code bots, AI agents that hand over to a person, integrations including HubSpot, Salesforce, Shopify and Zapier.
Twilio: a developer platform. More flexible, but you build the logic.

Two rules shape everything. People must opt in before you message them. And when a customer messages you, a 24-hour window opens where you can reply freely; outside it, only pre-approved templates. For inbound leads that's fine: they message, the window is open, the automation replies.

3. An AI step so replies don't sound canned

An automation platform sits in the middle (n8n, Make or Zapier) and calls a model like GPT or Claude. The prompt carries your services, your tone, what the AI may and may not say, and the lead's message. Out comes a short reply plus a few labels: service wanted, urgency, whether a human should take over.

trigger  : new form entry / email / WhatsApp message / CRM record
ai       : read message → draft reply → label lead
check    : plain rules block prices, promises, anything off-limits
send     : email or WhatsApp, with a booking link
log      : lead + reply + labels to CRM, alert owner if needed

If you want lead data to stay on your own servers, self-hosted n8n is the option there.

4. A booking link

A booking link turns a reply into a next step.

Calendly: routing forms can qualify a lead and send them to the right person's calendar, with routing from HubSpot, Marketo and Pardot forms and Salesforce account lookup.
Cal.com: open source, self-hostable, with routing and round-robin for teams.

Both connect to Google and Microsoft calendars.

Most good setups combine three of these: a trigger from the CRM or inbox, an AI step for the wording, and a booking link at the end.

Before you switch it on

Speed is the easy part. These take more care:

Data: know where lead data goes and read each tool's data processing terms, especially the AI provider's. In India that means the DPDP Act, 2023; GDPR if you have European leads.
Consent: only message people on WhatsApp who opted in, and make opting out easy.
Tone: run it on your last 20 real enquiries and read the replies out loud. Fix the prompt or the rule when one sounds wrong.
Guardrails: no prices, promised dates or complaint handling on its own.
Handover: decide which leads go straight to a person, with the full conversation attached.
Monitoring: alert on failed runs so a broken connection doesn't silently drop leads.

If all you need are template replies, a CRM workflow and a booking link is an afternoon's work. AI-tailored replies across several channels, with guardrails and monitoring, take longer to get right.

Originally published on dizilus.com.

Automating invoice processing in 2026, from built-in features to custom flows

Henil Mehta — Sat, 10 Oct 2026 09:51:17 +0000

Invoice automation gets sold as one big product. In practice it's five small steps, and the tools differ in how many of them they cover:

Capture the invoice from an inbox, folder, phone photo or supplier portal.
Extract supplier, invoice number, dates, tax and totals, sometimes line items.
Check for duplicates, unknown suppliers and totals that don't add up.
Approve: send anything unusual to a person.
Post the bill to your accounting software with the PDF attached.

Here's how the options stack up, cheapest first.

Start with what your accounting software already does

Xero: Hubdoc gives you an email address to forward bills to, extracts supplier, amount, invoice number and due date, and publishes to Xero with the document attached.
QuickBooks Online: set up a forwarding address ending in @assist.intuit.com. Forwarded bills are read and details like dates and amounts get filled in. Gmail users can auto-forward supplier emails to it.
Zoho Books: autoscan reads uploaded or emailed bills and pulls vendor, date, amount and line items. Each organisation gets an address for forwarding files into its Inbox.
TallyPrime: its strength is import, not reading. You can bring transactions in from Excel with a mapping template (Alt+O, then Import) or through its XML interface. The usual pattern is to extract elsewhere and import.

Built-in capture is the cheapest place to start. It's weakest on unusual layouts, long line-item tables and custom approval rules.

Dedicated AP tools

These go deeper than built-ins:

Dext Prepare: made for small businesses and their accountants. Email, app or upload in; supplier, totals, tax and line items out; duplicates flagged; publishes to Xero, QuickBooks, Sage and others.
Rossum: AI document processing for bigger finance teams. Handles complex line items and new layouts, plus GL and tax coding, PO matching, approvals and duplicate detection.
Nanonets: AI extraction with AP workflows on top: header fields, line items, PO matching, approval routing, connections to QuickBooks Online, Xero, NetSuite and SAP.

Worth it with steady, high volumes, several approvers or POs to match. For a few dozen invoices a month, probably more than you need.

Google and Microsoft document AI

If you want extraction inside your own workflow:

Google Document AI: the Invoice Parser returns invoice number, supplier, dates, amounts, tax, VAT details and line items. There's an Expense Parser for receipts and a Custom Extractor for your own fields.
Azure AI Document Intelligence: the prebuilt invoice model reads invoices, utility bills, sales and purchase orders from PDFs, scans and photos, in 27 languages, returning key fields and line items as JSON.

Neither posts anything to your books on its own. They suit custom approval processes, ERPs without good connectors, or strict rules on where data goes.

Custom flows in n8n, Make or Zapier

When your process doesn't fit a packaged tool (invoices arriving in several places, approval rules that depend on supplier or amount, data going to more than one system), build it. The shape is simple:

email trigger → AI reads the PDF (GPT, Claude, Gemini or Document AI)
             → rule-based checks
             → create bill in Xero / QuickBooks / Zoho Books
               (or write to a sheet for Tally to import)

n8n gives the most control, can be self-hosted so invoice data stays on your server, and has templates for emailed PDF invoices.
Make is good for visual flows with lots of branching, like different checks per supplier.
Zapier is quickest to set up, with AI by Zapier steps on paid plans.

Before you choose

Put your own invoices in front of you and check:

Where do invoices arrive today, and how many a month from how many suppliers?
Does the tool read line items or just totals? (Many charge more for line items.)
Does it handle your tax details, like GSTIN and HSN codes on Indian invoices, or VAT numbers?
Does it connect directly to your accounting software, or only via a spreadsheet?
How does it handle duplicates and unknown suppliers?
Where is the data stored, and who can see it?

The single most useful thing you can do: collect 20 to 30 real invoices, including the messy ones, and run every shortlisted tool on the same set. The differences show up fast.

The order we'd set it up in

Turn on the capture feature you already have and see how far it gets on real invoices.
If it struggles, add an AI step that reads invoices and runs your checks before anything reaches your books.
Send anything unusual to one person, with a note on what looked wrong.
Post clean invoices with the original attached, filed by supplier and month.
Review the flagged ones for the first few weeks and tighten the rules.

On clear digital PDFs, modern extraction is generally reliable. Blurry photos, handwriting and odd layouts are harder, which is why the check and review steps matter more than the extraction itself.

Originally published on dizilus.com.

Zapier vs Make vs n8n for small business automation in 2026

Henil Mehta — Sat, 10 Oct 2026 09:50:10 +0000

We build on all three and pick per client. All three handle most small business automations well. The differences show up in where they run, how hard they are to learn, and how they count usage.

Quick naming note, because it trips people up: Zapier calls a workflow a Zap (a trigger plus actions, built step by step), Make calls it a scenario (modules on a visual canvas), and n8n just calls it a workflow (nodes on a canvas).

Learning curve

Zapier is the easiest start. The editor is linear, and Copilot can draft a Zap from a plain-English description. Most people get a simple two-step Zap working in an afternoon.

Make takes a bit longer. The canvas shows every branch, filter and loop, which is great once it clicks. Routers, iterators and aggregators give you a lot of control, but they're new ideas for most people.

n8n is the steepest. It's visual, but it assumes you're comfortable with JSON, and it's easy to drop into a Code node for anything unusual. Developers tend to like it. Non-technical teams usually need help to start.

Judge this on the workflow you actually need, not a demo. Two steps are easy anywhere. Branches, loops and error handling are where the gap shows.

Self-hosting

This is the biggest practical difference: only n8n can be self-hosted.

Zapier is cloud-only.
Make is cloud-only too. It has an on-prem agent so scenarios can reach systems inside your network, but the platform runs in Make's cloud.
n8n runs on n8n Cloud or your own server. The self-hosted Community Edition is free for your own internal business use under the Sustainable Use License.

Self-hosting matters when client data has to stay on infrastructure you control. The trade-off: someone has to patch, back up and monitor that server.

Pricing: compare the unit, not the headline

	Counts	Notes
Zapier	tasks	A task is a successful action step. Triggers and built-ins like Filter, Formatter, Paths and Delay don't count.
Make	credits	Most module actions use one credit. Routers and error handlers don't.
n8n	executions	One full run of a workflow, however many steps it has.

So long, many-step workflows tend to be cheapest on n8n, short ones are fine anywhere, and high-volume multi-step flows are where Zapier's per-task model adds up fastest. Plans change often, so check the current pricing pages before deciding.

AI features

All three now treat AI as core:

Zapier: Copilot, AI by Zapier steps, Zapier Agents, Zapier MCP.
Make: Maia for building scenarios by conversation, Make AI Agents, and modules for OpenAI, Claude and Gemini.
n8n: an AI Agent node, nodes for chat models, memory and vector stores, MCP support, and an AI Assistant in the editor.

For everyday jobs like reading an invoice, classifying an email or drafting a reply, any of them works. n8n gives the most control over how an agent is wired. Zapier gets you to a working result fastest.

Same job, three platforms

A small agency wants every website enquiry added to its CRM, scored by AI and posted to Slack with a suggested reply.

On Zapier, a sales manager could build it in an afternoon.
On Make, the whole flow sits on one canvas, with a router sending hot and cold leads down different paths.
On n8n, it could run on the agency's own server, with an AI Agent node that checks past projects before writing the reply.

All three work. The right one depends on who maintains it and where the data is allowed to live.

How we'd pick

Zapier when your team will build and maintain it themselves, workflows are short, and you need a connector for a less common app.
Make when there are lots of branches, loops or data transformation and you want the whole flow on one screen.
n8n when data must stay on your servers, workflows run many steps per execution, or you want full control over AI agents.

Plenty of businesses end up on more than one. That's fine, as long as someone knows which automation lives where.

One thing people ask: can you move Zaps to Make or n8n? Not automatically. Each stores workflows in its own format, so it's a rebuild. The logic carries over, though, so the rebuild is faster than the original.

Originally published on dizilus.com.

I analyzed why I never finish online courses. Then I built the fix.

Henil Mehta — Tue, 12 May 2026 16:37:56 +0000

Last December I did a stupid little personal audit. I opened my Udemy library, my Coursera transcript, and the Pluralsight history, and I counted:

47 courses purchased or enrolled in over 6 years
3 finished
11 "in progress" (defined as: I'd watched between 10% and 50%)
33 untouched after lesson 2

A 6.4% finish rate.

I've been a software engineer for 9 years. I'm not lazy. I genuinely like learning. So why was I bleeding money and time into courses I never finished?

I spent the next two weeks doing root-cause analysis on every abandoned course in that list. Patterns showed up fast.

Pattern 1: Pacing mismatch

The single biggest predictor of abandonment was a mismatch between the course's pacing and my actual weekly availability.

Every course on Udemy and Coursera is built around an implicit assumption about how much time the learner has. A "10-hour Python course" assumes you watch it in roughly one or two sessions. A 12-week Coursera specialization assumes you can spend 5–7 hours per week on it for three months.

In practice, my real life has 4–7 hours per week for self-learning, randomly distributed, with occasional weeks of zero. Every course I bought was designed for a different distribution of time than the one I actually had. By week 3 I was always behind, and once you're behind, the calculus of "do I catch up or quit?" almost always favors quit.

Pattern 2: The illusion of choice

Half the courses I bought were never even started — I'd just buy three more during a sale and forget about them by Tuesday.

I now think this isn't a discipline problem. It's a UX problem. When the platform's primary surface is "here are 250,000 courses, good luck," the cognitive cost of choosing exceeds the cognitive cost of learning. So you do the thing your brain experiences as cheaper, which is keep browsing.

On HN this week someone called this "the Illusion of Choice." Perfect phrase. The marketplace business model is fundamentally at odds with the learner's outcome.

Pattern 3: No real-time unblock

The courses I did finish had something in common: I had a coworker, a Discord, or a tutor I could ping when I got stuck. The courses I abandoned, I got stuck around lesson 5–10 and just... stopped.

Pre-recorded video can't unblock you. The discussion forum on Udemy answers questions in 3–7 days, by which point you've moved on. Coursera's mentor model is better but only on paid specializations and still asynchronous.

Pattern 4: No catch-up logic

Miss a week on Coursera and the deadlines keep moving. You fall further behind. Eventually the dashboard becomes a guilt monument and you stop opening it.

Udemy doesn't have deadlines, but it also doesn't have any structure — so "I'll catch up next week" turns into "I'll catch up next month" turns into never.

What I built

I built Solohustller (https://solohustller.com) around these four findings, intentionally inverted:

1. Pacing matches your actual time. You declare your weekly hours up front (5h / 10h / 15h). The generated curriculum is sized to fit, not to fit "the average learner."

2. One course, not a catalog. You describe a goal in plain English ("learn React.js fundamentals", "Spanish conversational basics in 3 months") and the system generates exactly one curriculum. No browsing, no choosing between 47 versions.

3. AI tutor that knows your current lesson. When you're stuck, you don't ask a generic chatbot — you ask one that has the context of the exact lesson you're on and explains it in terms of what you've already covered.

4. Adaptive rescheduling. Miss a day, the next week rebalances. There's no guilt dashboard.

What I learned in the first 1,000 users

A few things surprised me:

People don't want fewer choices, they want zero choices. The most common feedback was "I love that I didn't have to pick." The cognitive relief of "the platform decided" was the actual product.
Quizzes are stickier than I expected. Weekly quizzes outperformed badges and streaks as a retention mechanic. Probably because quizzes give a binary "did I get this?" signal, which dopamine systems love.
98% completion isn't a flex, it's a baseline expectation. Once you've designed for actual completion, you can't go back to 10% being acceptable.
The hardest unsolved problem is cold-start curriculum quality. The first 90 seconds of the generated course decides whether the user trusts the system. Still iterating.

Where this might be wrong

Two honest concerns about the model:

It works less well for niche/expert topics. A generated curriculum for "Idiomatic Rust async patterns" is going to be worse than a great instructor. Marketplaces still win for niche depth.

It works less well for branded credentials. A Coursera certificate from Stanford has resume signal that a Solohustller certificate doesn't. If you need the brand, use the brand.

But for the 80% of self-learning that isn't niche-expert or credential-driven — "learn this skill, fit it into my life, actually finish" — I think the marketplace model is the wrong shape for the problem.

What I'd love to hear from this community

If you've been the person with 30 abandoned Udemy courses, what would you have wanted instead?

If you're a builder, what did I miss in the root-cause analysis?

And if you want to try Solohustller, the free tier covers Week 1–2 of any course and 50 AI tutor queries — no credit card. Would value brutal feedback on the homepage and the first 90 seconds of onboarding specifically.