Blog · AI agent basics
Why does my AI agent say “I don’t have that information”?
Your AI agent says “I don’t have that information” when the answer isn’t in its knowledge base or instructions. That is the safe behavior: it is better than making something up. The problem is that most platforms don’t treat the deflection as a failure, so nobody finds out, and the same question fails again for weeks. The fix is to find those moments in your transcripts, group them into a short list, get each answer from the person who knows it, and add it to the agent.
What is actually happening?
A customer-facing AI agent answers from what it has been given: its knowledge base, its instructions, and any tools it can call. When a guest asks something outside all three, a well-built agent declines instead of guessing. You hear “I’m not sure, please call the front desk” or “I don’t have that information.”
The conversation then ends politely. On most dashboards it looks like a normal, completed conversation.
Why doesn’t anyone notice?
Agent platforms log every conversation, and many score them for things like completion or sentiment. A score tells you how a conversation went. It doesn’t tell you what the agent didn’t know.
At real volume, nobody reads transcripts one by one. So the same missing fact can be asked dozens of times a week, in different words, and never show up as a problem.
Are all of these misses the same?
No. They fall into three groups, and each needs a different fix:
- ENTRY: a fact is missing, such as a pet fee or whether there is EV charging. One reply from the owner fixes it.
- TOOL: the agent needs a live lookup or an action, such as checking room availability for specific dates. No knowledge-base entry can fix it; it needs an integration.
- DEMAND: the customer wants something you don’t offer. There is nothing to fix in the agent, but the owner should know how often it is asked.
How big is the TOOL share?
It can be large. In a public run on the MultiWOZ 2.2 research corpus (10,437 dialogues), 41% of the unanswered moments in the 62 named gaps could be fixed with one knowledge entry, and 59% needed a tool. That corpus is booking dialogue with a bounded database, so its split is not a prediction for any one business. It does show why counting misses is not enough: you have to sort them before you can fix them.
How do I find every time my agent couldn’t answer?
- Export a week of transcripts from your platform.
- Search for deflection phrases: “I don’t have that information”, “I’m not sure”, “please call”, “I can’t help with that”, “let me transfer you”.
- For each hit, copy the customer’s question, not the agent’s reply. The question is what tells you what was missing.
- Group questions that one answer would cover. “Is there a charger for my Tesla?” and “Do you have EV charging?” are the same gap.
- Rank the groups by how often they were asked, and label each one ENTRY, TOOL or DEMAND.
How do I fix it?
Start with the most-asked ENTRY gaps. Ask the person who actually knows, usually a shift manager or the owner, one question at a time, and add the answer to the agent’s knowledge base. Send TOOL gaps to whoever owns integrations. Share DEMAND gaps with the owner as a count.
Then repeat every week. New questions appear with seasons, events and changes to the business, so a knowledge base that was complete in spring is out of date by summer.
Common questions
Is it bad that my AI agent says it doesn’t know?
No. Declining is safer than inventing an answer. The problem is when the same question keeps failing because nobody adds the answer.
Can my agent platform find these for me?
Platforms log and score conversations, and some group topics. The harder part is getting each answer from the person who knows it and putting it back into the agent, across every platform you use.
How often should I review unanswered questions?
Weekly is a good default. Most gaps repeat within days, and new ones appear with seasons, events and policy changes.
Connect read-only, or let us run a free 48-hour gap report for you.