How to train AI on your business: tone, rules, limits.

The short answer

You train AI on your business by writing down how you want the work done: your tone, your rules, and the things it must never do. Short sentences a human can read and a helper can follow. We call them laws, and every helper has to follow them to prove its work.

The generic way is useless

An AI that answers the generic way sounds like every other business on the internet. Warm where you're formal, exclamation marks where you'd never use them, promises you'd never make. Your customers can tell. A helper that does the task the generic way is useless, even when the facts are right.

Write laws, not prompts

Prompts are wishes. Laws are requirements. A prompt says "be friendly." A law says "replies sound like us: warm, short, no exclamation marks." The difference shows up on the day the AI is busy, or confused, or wrong.

What a law looks like

• "Never promise a delivery date without checking the calendar first."
• "Refunds over 200 go to a human, always."
• "Sign every customer email with the first name of a real person on the team."
• "If you're not sure, say so and hand over. Never guess."

If you can say it, it can be written as a law. And if a helper breaks a law, that counts as a failed proof, and a human hears about it.

Mistakes become laws

The best laws come from real mistakes. When a helper does something you dislike, that dislike becomes a law, and it never happens twice. Businesses change their minds too: tell us in plain words ("stop offering the discount on Fridays") and the law updates everywhere the same day.

You're not training a model. You're writing down how your business works, once, so every helper does it your way.

Further reading

• Your way of working: how laws get written and changed.
• Show done, never say done: how following the law gets checked.