The question we hear most from business owners — at events, in emails, on the first call — is some version of this: "What if the AI tells my customer something that isn't true?" A price you never set. A refund you never offered. An opening time you don't have.

It's the right question. Here's the honest answer: yes, AI tools make things up — and no, you don't have to accept that. The fix isn't a smarter model. It's boundaries.

Why it happens

These tools are built to always produce an answer. When they know, they answer from knowledge; when they don't, they complete the pattern anyway — fluently, confidently, and sometimes wrongly. It isn't lying; there's no intent. But your customer can't tell the difference between a confident answer and a correct one, and neither can the tool.

Understanding this changes how you fix it. You will never instruct a general-purpose AI into perfection. What you can do is take away its opportunities to improvise.

Boundary one: it answers from your documents, not its imagination

Give the AI your actual material — your services, prices, policies, FAQs — and restrict it to answering from that material only. Most business questions aren't creative questions; "how much is a haircut" has exactly one right answer, and it's in your price list, not in the model's memory.

Two practical notes. First, this is not "training" — you're not teaching a model; you're handing over reference documents, and you can start with a few pages. Second, consistency beats volume: two documents that disagree about your prices will produce confident nonsense faster than any model flaw. Fix the documents before you blame the AI.

Boundary two: give it permission to not know

An AI that must always answer will eventually invent one. So take that obligation away — explicitly. The single most effective line you can put in any business AI setup is some version of:

"If you're not sure, or the answer isn't in the provided material, say you don't know and pass the conversation to a person. Never guess."

An "I'll check with the team and get back to you" costs you thirty seconds. An invented discount costs you money and trust. Make not-knowing an approved outcome, and most of the hallucination problem disappears with it.

Boundary three: some things are never the AI's to say

Even with the first two boundaries, draw hard lines. Prices, refunds, discounts, exceptions, promises, availability commitments — anything that binds your business — should either come from your documents verbatim or from a human, never composed by the model. This is the same rule we apply in systems we build and run ourselves: the AI handles the volume, and anything with authority in it goes to a person, with context.

What about our data — is the AI learning from it?

The mirror-image worry, and worth answering while we're here: on the business tiers of the major tools, your data isn't used to train the models by default, and giving an AI your documents as reference material doesn't teach the model anything permanent. It's a knowledge base, not a lesson. Check the data controls when you sign up — and if a vendor can't answer "is our data used for training?" in one clear sentence, that tells you something too.

The part most businesses skip

Here's where this usually falls apart: the owner understands all of this, sets it up carefully in their own account — and meanwhile three staff members are pasting customer chats into three personal free accounts with none of these boundaries.

Boundaries only work when they're shared. That means writing them down: which tool is approved, what never gets entered, when the AI must hand over to a person, who checks its output. One page is enough — most businesses just never write the page.

That page is exactly what our free AI Usage Policy Starter drafts for you: answer a few questions about your business and get a practical starting policy you can edit and put in front of your team this week. It runs in your browser; nothing you enter is stored or sent anywhere.

The AI that makes things up is the AI nobody gave rules to. Rules are cheap. Write them.

Running the agent yourself? The practitioner version of this article: Anatomy of a business AI agent →