To train an AI agent for your business, you write things down. You do not retrain the AI. Give it a job description, a set of reference pages and a checklist for each task. Then write every correction you want kept into those pages, with who said it and the date. Most of the work is the corrections.
VantaSoft runs on a team of 20 AI agents (17 with company jobs, plus three personal assistants added on November 12), and every one of them learned the business this way. On November 13, 2026 we counted what that leaves behind. Our content agent's job description is now about 9,300 words long, and it names an owner and a date 85 times. Below is what goes on each page, how corrections get recorded, the counts for the whole team, and a sheet you can use for your own agent.
What does it mean to train an AI agent?
For a small business, training an AI agent means teaching it your business in writing. The AI underneath already reads and writes well. What it lacks is your prices, your customers, your rules and your way of doing things. The tools most owners meet all work this way: you write instructions and attach reference files.
In ChatGPT, a custom GPT has an instructions box and a place to upload files. OpenAI's own guide to building one says to "use knowledge for reference material, not rules or behavior," and to put "rules, tone, and workflow guidance in instructions" (Creating and editing GPTs, OpenAI Help Center). Claude's projects take uploaded documents plus project instructions (What are projects?, Claude Help Center). Microsoft Copilot Studio agents draw on knowledge sources such as SharePoint sites, uploaded documents and public websites (Knowledge sources summary, Microsoft Learn).
Our agents are built differently from those, but the split is the same. Everything they know about how VantaSoft works, they read from pages we wrote. When one of them gets something wrong, we change a page.
What do you write down to train an AI agent?
You write four kinds of pages. A job description says what the agent does and for whom. Reference pages hold the facts it looks up. A checklist covers each task it repeats. A record of corrections holds every instruction you want kept. Keep rules and facts apart: rules rarely change, while facts like prices change often.
| What you write | What goes in it | What ours looks like | Who keeps it current |
|---|---|---|---|
| A job description | The job, who it serves, what it does alone, what needs your approval, who it answers to | One per agent. A new one starts at about 1,300 to 1,800 words | The person the agent works for |
| Reference pages | Prices, policies, services and processes. One subject per page, dated | A shared library. Our knowledge agent checks up to eight pages a day against the live systems | Whoever owns that subject |
| A checklist for each task | The steps in order, plus one line for each mistake the agent has made | Our content agent's research checklist gets a dated line after each lesson | The agent, with a daily read by our knowledge agent |
| Corrections | Each instruction you want kept, in your words, with your name and the date | Written into the job description itself, so the agent reads it before every task | A person, or an agent whose job that is |
Here is how a checklist grows. On October 31, our content agent was pricing sales tools for an article. One vendor's page showed a single set of prices beside both "billed monthly" and "billed annually," and they turned out to be the annual prices. That day its research checklist gained one line: click each billing option and read both prices.
Our agents also keep a notebook of things they find out while working, such as which page has the current price list. We trust it least of all. When a note turns out to be wrong, it gets fixed that day. An agent will act on an old note as readily as on a new one.
The job description is the page most owners already know how to write. Ours follow the same outline as a hire's, and there is a blank to copy in our guide on how to hire an AI agent.
How do you correct an AI agent so the fix sticks?
Write the correction where the agent reads it at the start of every task, in your own words, with your name and the date. A correction you type into a chat can be gone the next time the agent starts work. The written one is there every time, and the date tells you later why it exists.
Until November 7, our owners had to tag a message with the word "remember" before a correction would be written into an agent's rules. Most corrections come mid-conversation, while someone is busy with something else. Our CEO asked for the recording to be taken off the owners' plates.
Now Walt, our knowledge agent, does it every morning. He reads what the owners said to each agent the day before and picks out the instructions meant to last. He writes each one into that agent's job description with the owner's words and the date, then tells the agent what changed. When he cannot tell whether a remark was meant to last, he asks the owner instead of guessing.
Small preferences go through the same path. On November 12 our CEO asked one agent to number its questions whenever it asks several at once, so he can answer by number. That is now in the rules every one of our agents reads.
Recording too much is the other way to get this wrong. In Walt's first week, an instruction our CEO had given an engineering agent back in July was written into that agent's rules. He had it taken out the same day, because it was not a rule he wanted added. Since then, anything an owner said more than a month ago goes back to that owner before it is written down.
The opposite mistake, an agent writing a rule nobody gave it, cost us six weeks of narrower articles. That story is the third one in what went wrong when we gave AI agents real work.
How many rules does a trained AI agent end up with?
More than most owners expect, and the number follows how closely the work gets read, not how long the agent has been around. We counted two things for each of our agents on November 13, 2026: how long its job description is, and how many times it names an owner and a date.
| The agent's job | Words in its job description | Times it names an owner and a date |
|---|---|---|
| Content and social media | about 9,300 | 85 |
| Building and running agents for our clients | about 9,600 | 61 |
| Business development | about 12,900 | 40 |
| Finance | about 2,700 | 36 |
| Email outreach | about 3,000 | 19 |
| Chief of staff | about 3,000 | 18 |
| Engineering | about 3,400 | 16 |
| Client projects | about 3,000 | 15 |
| Proposal editing, set up October 7 | about 1,500 | 5 |
| Compliance, set up November 12 | about 1,400 | 1 |
| Three personal assistants, set up November 12 | about 1,300 to 1,400 each | 2 to 5 |
We counted these automatically: each place where an owner's name sits within a few words of a date. Most of the hits are rules. A few record setup choices, such as who picked the agent's name.
The content agent tops the list because its work is public. The owners read what it publishes, and each comment that should hold for next time becomes a dated rule. The finance agent has a short job description and a lot of dates, likely because money attracts precise instructions. The proposal editor has worked for five weeks and collected five. Her edits go to a person for review before anything leaves the building, which may be why fewer of her mistakes have needed a standing rule.
Longer is not better. The longest job description on the team runs to nearly 13,000 words. In a page that long, a new rule can quietly contradict an old one. Walt trims one job description every morning: he cuts repeats and moves old history out to our reference library, and every rule stays.
How long does it take to train an AI agent?
The first version takes an afternoon of your time. The training goes on for as long as the business keeps changing. A new agent of ours starts with a job description of about 1,300 to 1,800 words. The ones that have done months of real work carry dozens of dated rules on top of that.
Expect most of your corrections in the first weeks, while the agent meets your real work for the first time. Read everything it produces in the first month. After that, the corrections slow down, and they mostly come when something in the business changes: a new price, a new service, a new person to route things to. Each of those is a page update, and it takes minutes.
If you are setting up the first version now, our guide on how to implement an AI agent covers what to gather before launch, including the rules that live in someone's head.
A correction sheet to keep for your own agent
Keep one row per correction. Once a month, read the sheet. Any correction you have given twice belongs in the job description, and any row marked "just this once" should stay out of it.
| Column | What to write | Example: an agent drafting replies for a small contractor |
|---|---|---|
| Date | When you corrected it | November 13 |
| What it did | The output, briefly | Told a customer the job would take two to three days |
| What you wanted | In your own words, the way you would say it to a new hire | "We never give a time on the first reply. Say we'll confirm dates when we book the visit." |
| From now on, or just this once? | Your call. If you are unsure, it is just this once | From now on |
| Where it goes | Job description, a reference page, a task checklist, or nowhere | Job description, under what it never promises |
| Who said it | Your name | The owner |
| Did the next one come out right? | Check the next three times the same thing comes up | Yes, all three |
The example row is illustrative, not from a customer. The last column matters most. A correction that did not change the next result usually landed on the wrong page, or it contradicts something already written. Look for the older line that says otherwise.
What gets in the way of training an AI agent?
Four habits undo most of the effort: correcting in chat and assuming it stuck, mixing facts into the rules so the price list goes stale inside the job description, letting the agent write its own rules, and never pruning until the rules start to argue with each other.
The first one is the most common. A correction typed into a conversation feels finished, and the agent agrees with you on the spot. The next morning it starts fresh from its pages.
Mixing facts into rules shows up later. A rule like "always quote the current hourly rate" stays true. A rule that says "our hourly rate is $95" is wrong the day you raise it, and nothing tells you. Keep the number on a dated reference page and have the rule point to it.
Stale reference pages do the most quiet damage. A page that is out of date is worse than a missing one, because the agent will act on it without a second thought. Give every reference page an owner and a date, and check the oldest ones first.
Frequently asked questions
Can I train my own AI agent?
Yes. If you can write a job description for a new hire and correct their work, you can train an agent. ChatGPT, Claude and Copilot Studio all take written instructions and reference files. Keeping those pages current as the business changes is the part that takes ongoing time.
Do I need to retrain the AI on my company's data?
Not for what this article covers. Your prices, policies and preferences change, and a written page can be updated in a minute. Every tool above reads its pages each time it works, so a change takes effect on the next task. OpenAI's guide also suggests tightening the instructions and adding examples before adding more tools.
How much does it cost to train an AI agent?
Mostly your time: an afternoon for the first version, then a few minutes per correction. If you have someone build and maintain the agent, the training is part of their work. Our builds start at $2,500 and include 16 implementation hours. Hosting and maintenance starts at $475 a month and includes small changes as your work evolves (see pricing).
Can an AI agent train itself?
Partly. Ours add lines to their own checklists and keep their own notes, and our knowledge agent reads every change the next morning. Rules are different. A rule counts only if a person gave it, and it carries that person's name and the date, so you can trace why the agent behaves the way it does.
Who in my business should train the agent?
The person who would have trained a human doing the same job. They know the edge cases and the customers. Name one owner for each agent, and route every correction through that person or through a record they read. Agents with several bosses and no record end up with rules that contradict each other.
Where to start
Open the instructions of the AI tool you already use and add the last three corrections you gave it, each with today's date and your name. Then start the correction sheet above. To see the jobs our agents do, read our AI agent examples. For deciding which actions still need your approval, see human in the loop for AI agents.
If you would rather have an agent built around your workflows and kept current for you, Start Automating. It starts with 30 minutes on your workflows.
