How to stop an AI assistant from inventing prices and promises
· 3 min read
Ask a language model about a product it doesn’t know and it will often answer anyway — confidently and plausibly. In a support chat that’s embarrassing. In sales it costs money: a price you can’t honour, a discount you never offered, a booking that doesn’t exist.
Why AI invents things
A language model is trained to produce a likely continuation of the text. When the facts are missing from its context, the most likely-sounding answer is still an answer. It has no built-in sense of “I don’t know this particular shop’s price for the 12 kg bag”.
Writing “never invent prices” in the prompt helps, but it isn’t a guarantee. A prompt is an instruction, not a lock: faced with an unusual question or a pushy customer, the model can still slip.
Rule 1: facts come from tools, not from memory
Items, prices, availability and free slots should come from a call to your real data — a catalog search, a calendar lookup, your shop’s API — every time. The model presents what the tool returned; it doesn’t recall it.
In Sales Core AI the agent searches your catalog or connected shop before recommending and shows the options with the exact prices the search returned. If the item isn’t there, it says so and offers what is.
Rule 2: an action happened only if the system confirms it
The riskiest sentence in an AI sales chat is “Done, you’re booked!” when nothing was booked. The fix is structural: the agent may say an order or a booking exists only after the tool that creates it returned success. Every call and its result is recorded, so you can check what really happened in each conversation.
Rule 3: hard limits live in code
Some rules are too important to leave to the model’s judgement. The system should check them at the moment of action, whatever was said in the conversation:
- A booking needs a contact and a time in the future.
- A discount above your policy limit is refused — the agent can offer to ask a manager instead.
- Cancelling a booking needs the customer’s explicit confirmation.
- Once a conversation is handed to a person, the agent stops taking actions in it.
Rule 4: “I don’t know” is a feature
When the knowledge doesn’t cover a question, the honest answer is “I don’t have that information — shall I ask a colleague?”. Then the question should go somewhere useful. In Sales Core AI unanswered questions land on a list in your dashboard; you answer once, and every agent knows it from then on.
Rule 5: test before customers do
Keep a set of test conversations with the tricky cases — the missing product, the discount hunter, the customer who wants a slot that’s already taken — and run them again after each change. A sandbox that never touches real orders or bookings makes that safe.
The short version: let the model talk, let the system decide. The model chooses what to say and which tool to call; your data and your rules decide what is true and what is allowed.