Best Practices· 6 min read

What To Do When Your Agent Keeps Saying "I Don't Know"

It is doing the right thing and it is still frustrating. Here is how to work out what is missing and fix it in the right order.


Seeing your own agent say "I don't have that information" is deflating. It is also the behaviour you want — the alternative is confident invention, which costs far more.

But a pattern of it means something specific: content is missing. Here is how to find out what, and fix it in the right order.

Get the list

Any decent product logs the questions it had nothing relevant for. Find that report — it is the most actionable thing a chatbot produces, because every line names a specific gap rather than a vague problem.

If you have to read transcripts instead, do that. It takes twenty minutes and works fine.

Sort by frequency, not by annoyance

The temptation is to fix whichever gap irritated you most. Resist it.

One question asked twelve times is worth more than nine questions asked once. The tail is long and mostly harmless; the head is where your customers are.

Then work out which kind of gap it is

Genuinely missing. The answer is not on your site anywhere. Write two sentences and re-crawl. Most gaps are this.

Present but buried. It exists inside a wall of marketing prose. Pull it out into its own section with a heading. Retrieval works on chunks, so a sentence surrounded by four irrelevant paragraphs retrieves badly.

Genuinely variable. "How much for a rewire?" has no fixed answer. Publish what it depends on: property size, occupancy, whether it is staged. That is honest, it is useful to the reader, and it lets the agent ask the right follow-ups instead of stalling.

Out of scope. Someone asked about something you do not do. The right fix is publishing what you do not do, so it can decline politely instead of going quiet.

The one to leave alone

Questions about a specific customer's account or order. Unless you have an integration, the agent cannot know, and it should not pretend to. Capturing the reference and handing over is the correct outcome.

What good looks like after a fortnight

Not zero unanswered questions — you will never get there, and chasing it means writing content nobody reads.

What you want is the head of the list handled, and the remainder being genuinely novel one-offs. When new conversations stop surfacing new gaps, you are done.

Check it is not a phrasing problem

Occasionally the content exists and the agent still misses it, because the page uses your internal vocabulary and the visitor used theirs.

Your site says "planned preventative maintenance". They typed "do you do servicing". Retrieval usually bridges that, but not always, and the fix is easy: add the customer's phrasing to the page. A sentence beginning "Also known as…" costs nothing and closes the gap for search as well.

Do not paper over it in the persona

There is a temptation to instruct the agent to "always give a helpful answer" when it hits a gap.

Do not. That instruction is precisely how you get invented prices and imagined policies — you have told it that saying nothing is unacceptable, so it produces something. An honest "I don't have that, shall I get someone?" is worth far more than a fluent guess, and it is the behaviour that keeps the rest of its answers trustworthy.

Related: what content your chatbot needs.

Frequently asked questions

Is 'I don't know' a failure?
It is the correct behaviour and much better than inventing something. But a pattern of it means content is missing, and that is worth fixing.
How do I find out what it could not answer?
Most products log unanswered questions. That list is the most actionable report a chatbot produces — every line is a specific gap.
Should I add content for every gap?
No. Work in frequency order. One question asked twelve times matters more than nine asked once.
What if the answer genuinely varies?
Publish what it depends on rather than a number. 'It depends on property size and access' lets the agent ask the right follow-ups.

Keep reading

What To Do When Your Agent Keeps Saying "I Don't Know" · SpideyChat