Why AI invents your prices, and the page that stops it

"Our standard maintenance package is KES 18,000 a month and includes two site visits."

Picture that sentence in an AI-drafted proposal a sales coordinator is about to send. It reads well. The number is plausible. But the package was renamed, repriced and restructured months ago, and the two site visits became one visit plus remote support. Nobody told the model, because nobody thought they needed to. It was asked to write a proposal, proposals have prices in them, so it supplied one.

Maybe the coordinator catches it. If your firm drafts with AI, it is worth asking how many sentences like that have already gone out without anyone catching them.

Why models do this

There is a clear explanation for this behaviour, and it comes from the people who build the models. In September 2025 OpenAI published a paper on why language models hallucinate. The short version: the way models are trained and scored rewards a confident guess over an honest "I don't know". A model that guesses gets some answers right. A model that admits uncertainty gets scored as wrong. So the guessing wins.

That matters for a business because your internal facts are exactly the things the model cannot know from its training. Your prices, your package names, your payment terms, your delivery windows. When the prompt needs one of them and the context does not supply it, the model does what it was trained to do. It produces the most likely-sounding value.

This is not limited to small firms with busy staff. In October 2025 Deloitte Australia agreed to partially refund the Australian government for a report that contained fabricated references and a made-up court quote. A large professional-services firm, with review processes, still shipped invented facts with its name on them.

The verified-facts page

The most effective fix we know for internal facts is unglamorous. Keep one page of the facts your AI is most likely to need, and have it read that page before it writes anything that leaves the building.

Each line on the page carries five things:

1. The fact. "Maintenance Plus, KES 15,500 per month, one site visit plus remote support."
2. Where it came from. The current rate card, a signed contract, the owner's decision.
3. The date it was last checked.
4. Who owns it, meaning who gets asked when it looks wrong.
5. Its status. Current, being changed, or retired.

Then add one rule the AI follows: if a fact it needs is not on the page, it says so and asks, instead of filling the gap. That rule does more work than any amount of prompt wording, because it turns the model's habit of guessing into a visible gap that a person can fill.

At Provecta we treat our own written facts this way. Our AI sessions read the company's context and standards before they start a task, and a missing fact is a reason to stop and ask, not a gap to fill with something reasonable.

What the rule looks like in use

Say a sales coordinator asks the AI for a proposal to a new client who wants maintenance on three sites. The facts page lists the Maintenance Plus price per site, but not whether there is a multi-site discount, because there is no standing one. Without the rule, the model would very likely invent a modest discount, since proposals for multiple sites often have one. With the rule, the draft comes back with the per-site price, the total, and a note at the top: "No multi-site discount is listed. Confirm whether one applies before sending."

That note is the whole point of the exercise. The coordinator now has a specific question to put to the sales lead, instead of a confident sentence to catch or miss. And if the sales lead decides that three or more sites get a set rate from now on, that decision goes onto the facts page as a new line, with a date and an owner, and the question never comes up again.

What it fixes and what it does not

It fixes invented internal specifics. Prices, names, terms, dates and figures that belong to your company stop being made up, because the model has the real ones in front of it.

It does not fix guesses about the outside world. If your AI is asked about a regulation, a court ruling, a competitor's pricing or market data, the facts page has nothing to offer, and the model will guess there exactly as before. That needs a separate check, a source you can open, or a person who knows.

It also creates a new failure if you neglect it. A facts page that nobody updates becomes a source of confident, repeatable wrong answers. The old price in that proposal would have been worse, not better, if it had been written on an official-looking page with a stamp of approval. The date column exists so that stale lines are visible.

And it does not remove the need to read what goes out. It reduces how much there is to fix. Someone still signs off.

Build the first version this week

Open a blank document and write the twenty facts your team quotes most often to customers. Prices and package names first, then payment terms, lead times, service areas and the names of the people customers deal with. Add a source, a date, an owner and a status to each line.

Then paste it at the top of the next five AI drafts your team writes, with the line "If a fact you need is not listed here, say so and ask." Compare those five drafts with the last five. You will see what changed without anyone having to measure it.

Check what wrong quotes and missed charges may be costing you at pgco.world/tools/revenue-leakage.