The AI did not hallucinate. Your documentation contradicts itself and nobody had noticed.

A support agent tells a customer they have 30 days to return an item. The contract says 60. The customer is furious, someone pulls the transcript, and the incident gets written up as an AI hallucination.

It was not. The help centre article says 30 days. It has said 30 days since 2019. The contract template was updated to 60 in 2023 and nobody went back to the help centre. Two humans had given the same wrong answer that month; nobody had escalated it, because when a human gets it wrong it is a mistake and when a machine gets it wrong it is a scandal.

This is extremely common

When you ground a system in a company's own material and require it to cite sources, you are running the first comprehensive consistency check that material has ever had. Every deployment turns up the same families of problem:

• The stale article nobody owns. Written by someone who left, still the top search result, wrong for two years.
• The policy that exists in three places. The help centre, the contract template and the internal wiki, all slightly different, all confidently maintained by different departments.
• The exception that lives only in someone's head. "Oh, we never charge restocking to trade accounts on terms." Written down nowhere. Everyone in the room knew.
• The account-level override in a system nobody mentioned. Negotiated freight terms, contract pricing, custom SLAs — in a field the discovery process never surfaced because it was obvious to the people who use it daily.
• The policy that contradicts the law. Rare, expensive, and worth finding.

Which is why "it said the wrong thing" is the wrong first question

When a grounded system produces a bad answer, there are three genuinely different causes and they need different fixes:

In our experience the third category is much the smallest, once retrieval is done properly. Most of what gets reported as hallucination is category one or two — the system faithfully reporting a mess that was already there.

The uncomfortable part

Companies find this genuinely unsettling, and it is worth naming before it happens rather than after. A grounding exercise will produce a list of places where your own written policy is wrong, out of date, or self-contradicting — and that list is longer than anyone expects. It is not a criticism of anyone. Every organisation over about fifty people has this, because documentation decays silently and nothing has ever forced a reconciliation.

Treat the list as an asset. It is the most valuable byproduct of the project, and several clients have told us afterwards that fixing the documentation was worth more than the automation. The AI just happened to be the first reader thorough enough to notice.