Why your AI agent answers wrong (and it isn't the model)

When a customer tells us their agent is “hallucinating,” nine times out of ten the LLM is doing exactly what we asked it to do — repeating what's in the knowledge base. The knowledge base is just wrong.

The four KB pathologies

Almost every bad answer traces back to one of these:

• Stale fact — “We're open until 7” (but you closed early last winter).
• Conflicting fact — Two policies say two different things; the agent picks one at random.
• Implied fact — Your team knows this from context; the doc doesn't say it.
• Buried fact — The right answer is on page 47 of a PDF; the retriever never finds it.

The 10-question audit

Pick the 10 questions your front desk hears most. For each one:

• Ask the agent in a sandbox call. Record the answer verbatim.
• Have a human who actually knows the answer grade it: correct, wrong, partial.
• For every wrong or partial: open the retrieved chunks. Identify which pathology applies.

You'll be done in under an hour and you'll have a list of KB edits worth more than any prompt tuning.

Write for retrieval, not for humans

Vector search retrieves chunks based on semantic similarity to the question. That means each meaningful fact needs to live in a chunk that reads like an answer to a likely question.

Bad: “Pricing tiers are detailed on the next page.”
Good: “Our standard cleaning starts at $189. Deep cleans start at $249.”

One source of truth

If your hours, prices, or policies live in three places, three versions will eventually disagree. Pick one canonical source (we usually recommend a single Google Doc per topic) and have everything else point at it. The agent reads the canonical version.

The “I'll find out” escape hatch

A good agent should be willing to say “I don't know. I'll find out and call you back.” This is far better than a confident wrong answer. Configure the agent's system prompt to default to this when retrieval confidence is below threshold. The dashboard logs every escape so you can backfill the KB.

What to clean first

• Hours of operation, including holidays and special closures.
• Pricing — especially anything that changed in the last 6 months.
• Service area boundaries and trip charges.
• Insurance / payment policies.
• Emergency / after-hours protocols.

Get those five right and 80% of your “hallucinations” go away.