Why a business agent must never invent prices

Ask a language model for a price it does not have and it will often give you one. Not because it is lying, but because producing a plausible answer is what it is built to do. For a business, a made-up price, delivery date or refund promise is not a quirk. It is a commitment your customer will quote back to you.

Why it happens

A model is trained to continue text in a convincing way. When the prompt asks "how much does a screen repair cost?" and no number is present, the most likely continuation is a number. Polite instructions like "do not guess" help, but they are a request, not a guarantee.

Three layers that work together

1. Ground it in data

Prices, stock and statuses come from a tool that reads your system. The model's job is to explain the result, not to produce it. When there is no tool result, there is no number.

2. Say what is not published

If your rates depend on scope, the agent should say exactly that and offer to pass the details to a person. A clear "I do not publish prices, he quotes once he understands the problem" is better than a vague guess and far better than a wrong figure. Write this into the agent's rules in plain words.

3. Check the answer before the customer sees it

This is the layer people skip. After the model writes its answer, run a deterministic check on it. It does not use another model, so it cannot be talked around. Mine looks for three things:

• Figures with no source. A currency amount that appears in no tool result and was not typed by the visitor is removed, with a note that pricing is not published.
• Links to nowhere. A link to a domain that appears in no source is stripped back to its text.
• False "done" claims. If the answer says a message was sent but the send tool never succeeded, a correction is added.

Test it on purpose

Add test cases that tempt the agent: "what is your hourly rate?", "how much for a mobile app?". The check is simple: the reply must not contain an unsourced figure. Run them after every change to a prompt or a model, because behaviour shifts when either changes.

Trust in an agent is built by what it refuses to say as much as by what it says.

The agent on this site works this way, and rates are not published here for exactly the reason above. If you want an agent that behaves like this in your business, see how an engagement works.