How to Train AI on Your Business So It Stops Guessing

How to train AI on your business is a question that sounds technical and is almost entirely editorial. There is no model to build and nothing to code. What decides whether an assistant is useful or embarrassing is one boring document: a written account of what you sell, what it costs, how you work, and what you say when things go wrong. Nearly every failure people blame on the technology is really a gap in that document, and the fix is a writing job you can start today.

Here is what goes in it.

Why a system invents things

When an assistant does not know something, it usually produces a plausible answer rather than admitting the gap. It will not sound uncertain. It will sound exactly as confident as when it is right.

That behaviour is not fixable by choosing a better tool. It is contained by two things: giving it the real answer so it does not have to guess, and keeping its brief narrow enough that it cannot wander into territory it has no information about.

A worker whose entire world is "answer questions about hours, location, and what we sell" cannot invent a returns policy, because it was never given room to. Narrowness is not a limitation you tolerate — it is the mechanism that makes the thing safe.

The document, section by section

Write it in plain language, as if explaining to a capable new employee on their first morning.

1. What you do, precisely

Not your positioning line. The actual list — services or products, what each one includes, what it does not include, and who each is for. Name the things you do not do, explicitly. This is what prevents the most embarrassing failure mode: confidently promising something you do not offer.

2. Prices, or the honest shape of them

Ranges, starting points, what moves the number up or down, what is always quoted individually. If you truly cannot publish numbers, write the rules — "jobs under X are quoted flat, anything involving Y is site visit first."

Rule: whatever is written here, the system may explain. Only a person may commit to it.

3. How the process actually works

Step by step, from first contact to finished. Timeframes. Who does what. What the customer needs to have ready. This is the section that removes the most hesitation, because people stall on not knowing what they are agreeing to.

4. The twenty questions you are asked most

Written as questions with complete answers. This is the highest-value section in the whole document and the one that also directly improves your website and your visibility in AI-generated answers — the same material does all three jobs.

5. Policies and edge cases

Cancellations, deposits, guarantees, what happens if something is late or wrong. Also the human ones: what you do when a customer is unhappy. Write your actual practice, not an idealised version.

6. Voice

Two or three examples of how you genuinely write to customers. Not style adjectives — real messages. This is what stops the output sounding like every other business in your category.

7. The escalation rules

Explicitly: what must always go to a person. Money commitments, complaints, anything unusual, anything about a specific customer's account.

How to train AI on your business without a big project

Do not attempt the whole document. It will not get finished.

Week one: write section 4 only — the twenty questions. An hour, using questions you already answer daily.

Week two: add sections 1 and 2. Another hour.

Week three: point one narrow job at it. Not everything — the after-hours questions, or the first reply to enquiries.

Week four: read what it produced and fix the document where it was wrong. Almost every bad answer traces to a missing or ambiguous line, and correcting the source fixes it permanently.

That loop — run, read, fix the document — is the entire method. There is no separate training step.

Where to keep it

One file, one place, plain text or a simple document. Not spread across five systems.

The reason is the maintenance problem: when a price changes, you want to edit one line and have everything reflect it. If the same fact lives in five places, it will be out of date within a month, and a confidently wrong answer is worse than no answer.

What not to put in

• Anything you would not want repeated. Assume everything in the document can end up in front of a customer.
• Customer personal data. Keep it out entirely.
• Passwords, keys, or account details. Ever.
• Speculation. If you are unsure of a fact, write "check with a person" rather than a guess. The document is allowed to say "escalate this."

How to keep it from going stale

Staleness is the real long-term risk, and it is invisible.

Put a recurring reminder — monthly is plenty — to reread two sections and fix what has changed. Attach it to something you already do, like invoicing day. And whenever a price, service, or policy changes, the document is part of the change, not a follow-up task.

A stale document is worse than none, because it makes wrong answers sound authoritative.

FAQ

Do I need special software to do this?

No. A single well-written document plus whatever assistant you already use covers most small businesses. Dedicated tools help at volume; they do not substitute for the writing.

How long should the document be?

However long it takes to answer the real questions without ambiguity. A few thousand words covers most small businesses. Length is not the goal — absence of gaps is.

Will it still get things wrong?

Occasionally, which is why escalation rules and the never-commit-money line exist. The goal is not perfection; it is that mistakes are small, caught, and fixable at the source.

Can I just point it at my website?

Partly, and it is a reasonable start — but most websites do not contain the operational detail: the edge cases, the policies, the real timeframes. Those are exactly what customers ask about.

Is this the same work as SEO?

Overlapping and usefully so. The twenty-questions section is precisely what makes pages rank and what AI assistants quote when somebody asks about businesses like yours. Write it once, use it three ways.

Want help writing it?

If an assistant has ever given a customer a vague or wrong answer about your business, the cause is almost certainly that the correct answer does not exist in writing anywhere.

We do this work: the document first, then the narrow workers, then the operations layer around them — and we will tell you plainly where a person still has to stay. If you want a straight read on what is missing, you can start it here.