How do you train AI on your own business?

Can you actually train AI on your own business?

No, and you do not need to. Nothing you type into ChatGPT, Claude or Gemini changes the model or teaches it about your company for good. What you can do is hand it your own information to read every time it answers. That is what project files and custom instructions are, and it works better than training would, because you can correct a wrong answer in a minute instead of a month.

The word training does a lot of damage here. It makes people picture a machine slowly learning their business over months of chatting, which is why so many owners keep typing the same background into a fresh window and wondering why it never sticks.

Here is the real shape of it. The model already knows how to write, summarise, argue and explain. What it does not know is you. Every tool has therefore added the same thing: a place to put your own material so it gets read before the answer comes out. OpenAI, Anthropic and Google all document a version of it, and they built it because it is the part that was missing.

There is a genuine training process called fine tuning, and it exists for a different job. It shapes how a model behaves, its format and its tone. Researchers who compared the two approaches head to head found that retrieval, meaning giving the model the material to read, beat fine tuning by a large margin on the least popular facts. Your prices in a town of eight thousand people are about as unpopular as a fact gets, so the cheap approach is also the one that works.

You are not teaching it. You are handing it the binder before it answers the phone.

Why does AI give such generic answers about my business?

Because it is answering about businesses like yours instead of about yours. It has read the whole internet on your industry and nothing at all about your prices, your process, or the four things you refuse to do. Given no information it fills the gap with the average, and the average is what makes the answer feel like it came from a brochure.

Watch what happens the first time you paste real details in. Ask it to write a quote reply with no context and you get a polite paragraph that could belong to anyone. Give it your actual prices, your deposit terms and two lines you have really sent to customers, and the same question produces something you could send.

Nothing about the tool changed in between. The only thing that changed is that it stopped guessing.

This is also why the first week with AI feels magic and the second week feels disappointing. The novelty carries the generic answers for a while. Then you start needing it to be right about your business, and it has nothing to be right from.

What is a business knowledge base, in plain terms?

It is what your business knows, written down in one place, in a form both a person and an AI tool can read. Prices, process, policies, the answers you give over and over, and the way you talk. I call it a brain because that is the honest description: the part of the business that currently lives in one head and nowhere else.

Most small businesses have this material already. It is just scattered. Some of it is in the owner's head, some in a quoting spreadsheet, some in six hundred sent emails, and a bit of it is on the website. That scatter is fine for a person who has been doing the job for ten years. It is useless to a machine, and it is useless to a new hire on their first Tuesday.

Two things make it more than a tidy document. It is written so the answer is findable rather than buried in a paragraph about your values. And it is complete enough that the tool never has to guess, because a tool that has to guess will, cheerfully, and in your name.

Worth separating this from a job it is often confused with. A knowledge base is about AI working for you. Getting named when a stranger asks an assistant for a plumber is a different problem, and I wrote about that one separately in how to show up in ChatGPT and AI search. Same technology, opposite direction.

What goes inside it?

Six things, and none of them are clever. What you sell and what it costs, how you actually work, the questions you answer repeatedly, who you are and are not for, how you talk, and the boring facts. Write those down and you have covered most of what any tool will ever need to answer as your business.

1. 01

   What you sell, and what it costs

   Every service or product, what is included in each, and the real price. Not a range you would be embarrassed by, the number you actually charge. This is the single thing AI gets wrong most often about a business, because it has no way to know and it will guess rather than say nothing.

2. 02

   How you actually work

   What happens after someone says yes. Deposits, timelines, what you need from the customer, what happens when they are late, how you handle a rush job. The stuff you explain on the phone twice a week without noticing.

3. 03

   The questions you answer over and over

   Write the question the way a customer asks it, then your real answer underneath. Twenty of these is a good start. If you have an inbox, you already wrote them, you just wrote each one once and then lost it.

4. 04

   Who you are for, and who you are not for

   The jobs you want and the jobs you turn down. This one is worth more than it looks, because an assistant that knows what you refuse stops writing you into work you hate.

5. 05

   How you talk

   A short note on tone, plus a few sentences you have actually written. Nothing makes AI output sound like a stranger faster than having no sample of the owner speaking.

6. 06

   The boring facts

   Hours, service area, phone number, address, licence and insurance details if customers ask about them, the year you started. Small, dull, and the reason half of a machine answer goes wrong.

The one people skip is how you talk. It feels like the soft item on the list and it is the reason output either sounds like you or sounds like a stranger wearing your name. Three or four real sentences you have actually written will do more than a paragraph describing your tone.

The one people get wrong is prices. There is a strong pull toward writing a comfortable range instead of the number, and a range teaches the tool to be vague exactly where a customer wants certainty. If a price depends on the job, write down what it depends on, which is more useful than either the range or a single made up figure.

Where does it live once it is written?

In whichever tool you use, in the slot every one of them now has. ChatGPT has Projects with uploaded files and instructions. Claude has project knowledge you upload documents into. Google has Gems that take files and custom instructions. The names differ and the shape is identical, which tells you this pattern is not a fashion.

ChatGPT

Projects hold uploaded reference files and a set of instructions that apply to every chat inside them. You add the material once and the conversations in that project start from it.

Claude

Projects have what Anthropic calls a knowledge base. You upload documents, text, code, or other files, and Claude uses them as the background for chats in that project. On paid plans it switches to a retrieval mode as the material grows.

Gemini

Gems are saved assistants with your own instructions, and Google documents adding files to give a Gem more context or to reference specific documents in your chats with it.

Everything else

Notebooks, agents, custom assistants, whatever arrives next quarter. The names keep changing and the shape does not: a slot for your instructions, and a slot for your files.

Because the shape is the same everywhere, the same document works in all of them. That is the quiet advantage of writing it as plain files rather than building it inside one company's product. You are not betting on which assistant wins.

Why does it have to be your file, not their feature?

Because the tools are rented and the knowledge is yours. Every AI product will happily hold your business information inside its own account, and that is convenient right up to the month the price doubles, the feature changes, or something better arrives. If your knowledge lives as ordinary files in your own folder, moving is an upload. If it lives only inside a product, moving is a rebuild.

I hold the same position on website platforms, and it comes from watching people get stuck rather than from any dislike of the companies. A tool you can walk away from is a tool you can negotiate with.

The practical version is small. Keep the master copy in your own storage, in a normal format you could open in ten years. Upload copies wherever you need them. When the tool changes, and it will, you are moving a file instead of reconstructing a decade of knowledge from memory.

The format I recommend, and the one I build in, is a set of markdown files kept in Obsidian. Markdown is plain text with light formatting, so the files open in any editor on any computer forever. Obsidian sits on top of that folder and gives you search, links between pages, and a map of how the pieces connect, which is what turns twenty documents into something you can find an answer in. The app is free, including for business use. The company asks organisations to buy an optional commercial licence at $50 US per user per year, and I mention it now so nobody feels ambushed later.

The part that matters is what happens if you stop liking Obsidian. Nothing. A vault is a folder, the notes are text files, and you keep working in Notepad. That is the whole reason I use it rather than a knowledge base product with an account and a monthly bill.

• Every price in it is the price you charge today, because a stale knowledge base is a confident source of wrong answers about your own business.
• It is written in plain sentences a person could read out loud, since anything a model has to decode it can also misread.
• It lives as ordinary files in your own folder, so moving to a different AI tool is an upload rather than a rebuild.
• It says what you do not do, not only what you do.
• It has no customer names, card numbers, health details, or anything else that would be a problem in the wrong hands.
• Somebody owns keeping it current, and that somebody has a reason to open it, such as the week a price changes.

How do you build one this weekend?

Open one document, write the eight sections below in order, then test it by asking the tool three questions you already know the answers to. The writing takes a weekend for a small business. The testing takes twenty minutes and is the step that turns it from a document into something you trust.

1. 01Open one plain document. Not a folder of twelve, not a spreadsheet. One document you can read start to finish. If you want the linked version from the start, install Obsidian first and make a new vault, which takes about two minutes.
2. 02Write what the business does, where, and for whom, in one paragraph. Say it the way you would say it to a stranger at a counter.
3. 03List every service or product with its real price and what is included. If a price depends on the job, say what it depends on.
4. 04Write what happens after someone says yes, step by step, including deposits and timelines.
5. 05Go into your sent email and pull out the twenty questions you answer most. Paste the question, then write the answer properly, once.
6. 06Add hours, service area, contact details, and the jobs you turn down.
7. 07Read it out loud. Anything you would not say to a customer comes out, and anything you had to explain twice goes in.
8. 08Upload it to whichever AI tool you use, ask it three questions you already know the answers to, and fix whatever it gets wrong.

Two habits decide whether it survives. Put it somewhere you will actually open, beside the files you use weekly rather than in a folder named admin. And tie the update to an event instead of a date, because nobody keeps a calendar reminder to review a document. When a price changes, the document changes in the same hour. That single rule is the difference between a knowledge base and a museum.

If you want a sense of what plain writing for machines looks like before you start, mine is at kootenaymade.ca/llms.txt. Different job, same register, and there is nothing clever in it on purpose. The related file convention is explained in my guide to llms.txt if you want the difference between the two.

When is it worth paying someone to do it?

When the knowledge is spread across more than one head or more than a handful of places, and when the reason it is not written down is that nobody has four spare days. Doing it yourself is fine and I have told people to. What you are buying from me is the interview that pulls it out of you, and the structure that makes the answer findable instead of buried.

The Brain is $800 for a solo or small business and from $3,500 for a team, confirmed on a short call before any work starts. The tiers are not about staff count. They are about how many places your business keeps what it knows, because every extra place is another gather and structure pass. A business with several locations, in a regulated field, or needing its systems joined up rather than read, gets quoted by scope instead.

It is a build rather than a subscription. You get plain files in your own storage, a session so you can keep adding to it, and no monthly fee for a document that is not changing. If you want me keeping it current as the business moves, that rides an ongoing support plan you may already have.

If you are not sure whether this is your bottleneck at all, say so and I will tell you. The AI and systems assessment is $300, credited in full against a build, and three of its five headings are about what to cancel and what to leave alone. Half the businesses I look at need a clearer website before they need anything with AI in the name, and what an AI business setup actually looks like covers the rest of the tooling. Full details of the build itself are on the Brain service page.

Sources and further reading

• Anthropic: what are Projects

  Anthropic documents uploading documents, text, code, or other files to a project knowledge base, and Claude using that material as background for the chats inside the project.

• Google: create a custom Gem in Gemini

  Google documents adding files to give a Gem more context or to reference specific documents in your chats with it, alongside custom instructions.

• OpenAI: model optimization guide

  OpenAI's own guidance on getting a model to do what you want starts with including the content the model needs to answer from outside its training data, before any talk of training anything.

• Soudani, Kanoulas and Hasibi: fine tuning vs retrieval augmented generation for less popular knowledge

  Peer-reviewed comparison published at the 2024 ACM SIGIR conference. The finding that matters to a small business: retrieval beat fine tuning by a large margin for the least popular facts, and a local business is the definition of an unpopular fact.

• Obsidian: pricing and licences

  The app is free with no sign up, including for commercial use. The commercial licence at $50 US per user per year is described on the same page as optional and encouraged rather than required, which is worth reading yourself before anyone tells you otherwise.

• Kootenay Made Digital: my own machine-readable summary

  A different file for a different job, and a good look at how plain this kind of writing should be. This one is aimed at search engines rather than at an assistant doing your work.

Frequently asked questions