A step by step guide for a business without an engineering team: which job to automate first, how to connect it, what it actually costs, and how to know when it is wrong.
Most guidance on AI agents is written for companies with an engineering team. This is written for a business that does not have one, and for anyone building their first agent before they build a complicated one.
It is not theory. We run an agent pipeline every day that reads company filings, scores what matters, writes it up and publishes it without anyone pressing a button. Everything below, including the failure modes and the costs, comes from operating that system rather than from reading about agents.
A caution before the steps. If you are automating something regulated, something that moves money without review, or something where a wrong answer harms a person, this guide is the wrong starting point and you should get proper help. What follows is how to build a first agent safely, not how to build every agent.
All prices and console paths below were checked on 13 August 2026 against primary documentation. Both move, so verify before you rely on them.
Step 1: write the job down by hand
Do the task manually and write down every step, including the ones you do without thinking.
A real example, chasing an unpaid invoice:
- 1. Open the accounting system, filter to invoices over 14 days old
- 2. Skip anyone who has replied in the last week
- 3. Skip anyone on a payment plan
- 4. Draft a polite reminder using last month's wording
- 5. Soften the tone if the client is a long-standing one
- 6. Send, and note the date
Steps 2, 3 and 5 are the ones people leave out, and they are exactly where an agent will embarrass you. If you cannot write the rule down, the agent cannot follow it.
You are done with this step when a competent new employee could do the job from your notes without asking you anything.
Step 2: decide whether you need an API key at all
Two routes, and most first agents should try the cheaper one first.
Route A, a subscription that runs tasks by itself. Claude's Scheduled Tasks and Routines run a saved prompt on a cadence, hourly, daily or weekly, with access to your connected tools. If your job is "every morning, read these emails and draft replies into a document", this is one subscription and no API key at all. Try this before anything else.
Route B, an API key plus an automation tool. Needed when the job crosses systems the subscription cannot reach, or when you want the automation to run on its own schedule independent of any app.
Route B is what the rest of this guide covers, because it is the one people get stuck on.
Step 3: get a key, and understand how you are billed
Anthropic: sign in at platform.claude.com, then API Keys, then Create Key. It begins sk-ant-. You see it once.
OpenAI: platform.openai.com, API keys, Create new secret key.
Understand the billing model, because it is not what most people assume.
The Claude API runs on prepaid credits. You load credits and calls draw them down. When they run out, calls fail. That is a harder guarantee than a monthly cap: you cannot be billed for more than you have already put in. Load twenty dollars, and twenty dollars is your maximum exposure.
The Spend Limits API you may read about is Claude Enterprise only and is not what a small business uses. Ignore it.
Either way: treat the key like a credit card number. Never in an email, never in a shared document, never in anything a customer could see. In Zapier or Make you paste it once into the connection and the tool stores it.
Step 4: build the smallest possible version
Three connections. Where it reads, the AI, where it writes.
In Zapier the vocabulary is Trigger, Action, Action. In Make it is modules joined by lines. The shape is identical.
Read. Start with a trigger you can fire on demand while testing. "New email in a specific Gmail label" is ideal, because you can drag an email into that label whenever you want to test.
Think. Add the AI step, search for the Anthropic or OpenAI app, paste your key, and choose the plain "send prompt" action.
Your prompt should say what to do, what to output, and what to do when unsure:
- You are drafting a payment reminder.
- Invoice data: (the data from your read step)
- Write a polite two paragraph reminder.
- If anything is missing or ambiguous, reply exactly: NEEDS HUMAN
That last line matters more than the rest. An agent with no way to say "I do not know" will invent something instead.
Write. Send the result somewhere harmless. A Google Doc, a draft folder, a Slack channel. Not the customer. Not on day one, not on day ten.
You are done when you can drag a test email into the label and see a draft appear somewhere only you can see.
Step 5: pick the cheapest model that can do the job
This is where most of the money is saved, and it is a one-line change.
Published Claude API prices, per million tokens, as of 13 August 2026:
- Claude Haiku 4.5: $1 input, $5 output
- Claude Sonnet 5: $2 input, $10 output
- Claude Opus 5: $5 input, $25 output
A token is roughly four characters, or about three quarters of a word.
What that means in practice. A payment reminder might send 1,500 tokens in (the invoice data and your instructions) and generate 500 tokens out. Three hundred of those a month:
- On Haiku 4.5: $1.20 a month
- On Sonnet 5: $2.40 a month
- On Opus 5: $6.00 a month
Scale it up to a thousand jobs a month with more context, say 3,000 in and 800 out, and Haiku is $7.00 while Sonnet is $14.00.
Drafting from a template is a Haiku job. Reserve the expensive models for work that genuinely needs judgement. Most first agents are running a model ten times more expensive than the task requires.
Two further discounts worth knowing. Prompt caching charges a cache read at 10% of the input price, which pays for itself after one hit, so a long fixed instruction block gets very cheap on repeat runs. Batch processing is 50% off both input and output if the work is not time sensitive.
Step 6: cap the run, not just the account
Prepaid credits stop a disaster. These stop the thing that causes one.
Limit the answer length. Every AI step has a max tokens setting. A two paragraph email needs about 500. Set it. Left unset, one confused run can generate pages, and output tokens are the expensive half.
Limit the steps per run. If your automation can loop or retry, cap the attempts. In our own pipeline the agent stops after a fixed number of tool calls and is told to produce its final answer with whatever it has. Without that, an agent that gets confused keeps trying, and every attempt is billed.
Limit the runs per hour. Most automation tools let you set this. If your trigger ever fires in a loop, this is what saves you.
Three ceilings: how long one answer can be, how many steps one run can take, how many runs can happen. Miss any one and the other two will not save you.
Step 7: check every output for a week
Let it run and read every result. Not a sample. Every one.
Keep a tally of correct, wrong, and needed a human. After a week you have a number instead of a feeling.
What to look for, in order of how often it bites:
Confidently wrong. The output looks perfect and the facts are wrong. This is the dangerous one, because nothing about it looks like a failure.
Quietly incomplete. It answers, but drops a case your notes covered.
Right but wrong tone. Fine for internal drafts, not for customers.
We publish a daily video built by our own pipeline. It once produced a technically perfect clip with the wrong person's face in it. Every automated check passed: the transcript matched the script, the file was the right size, the duration was correct. The only thing that caught it was a human looking at the picture.
Design the check before the agent exists. Decide what obviously wrong looks like for your job, and make that check automatic if you can.
Step 8: let it act, narrowly
Only now does the agent do anything real, and only for the cases it got right every time last week.
If it drafted 40 reminders and got 38 right, do not switch on all 40. Find what the two had in common, a missing field, an unusual client, a foreign currency, and exclude that case. Route it to you instead.
Add a rule that anything containing NEEDS HUMAN never sends automatically.
Keep the human review permanently, at a lower rate. Weekly instead of daily. Every model update changes behaviour, and you want to find that in a spot check rather than in a customer complaint.
What it actually costs
- The AI: for a first agent, single dollars a month, not tens. See the table above.
- The automation tool: free tiers cover a first agent. Paid plans start around twenty to thirty dollars a month.
- The real cost: the checking. A few hours a week initially, less later. It never reaches zero.
If someone quotes you thousands a month for a single-task agent, ask what in that number is not the three items above.
What not to do first
- Anything that moves money without a human approving it
- Anything a customer sees unreviewed
- Anything you cannot check in ten seconds
- Anything with health, legal or safety consequences
- Replacing a person before running the agent alongside them for a month
The one-page version
- 1. Write the job down until a new employee could do it
- 2. Check whether your existing subscription can already run it on a schedule
- 3. If not, get a key and load a small amount of prepaid credit
- 4. Read, think, write. Write to a draft, never the live thing
- 5. Use the cheapest model that can do the job, usually Haiku
- 6. Cap the answer length, the steps per run, and the runs per hour
- 7. Check every output for a week and count what went wrong
- 8. Let it act only on the cases it never got wrong, and keep checking
This is what has worked for us running a daily automated pipeline. It is not advice for your specific business, and anything touching money, customers or compliance deserves more care than an article can give.
This piece was originally published on Present of AI, where we cover what AI is actually doing in the world, no hype. Read more or get it in your inbox.
Top comments (0)