Why "I don't know" is a better answer than a guess

The moment it looks like a failure

If you're testing a chatbot for the first time, this answer can feel like a letdown. You asked a normal question and got sent somewhere else instead of a straight answer. It's tempting to mark that down as a miss, the same way you'd mark down a search engine that returned nothing. But that reaction gets the situation backwards. The chatbot just told you something true: it doesn't know, and it's not going to pretend otherwise. That's a much better outcome than it looks like in the moment.

What a guess actually costs you

Picture the alternative. Same question, but instead of admitting the gap, the chatbot answers directly: "Yes, all plans include a dedicated account manager who reaches out within your first week." That sounds like a win. You got a real answer, right away, no detour to sales. Except the chatbot didn't actually know that. It generated something that sounded plausible, and you have no way to tell the difference between an answer it verified and one it made up, because both come out in the same confident tone.

Nothing about a wrong guess announces itself as wrong. You act on it the same way you'd act on something true, because from where you're sitting it reads the same either way. The plain "I don't know" costs you a few extra seconds right now, and you know exactly what you're dealing with. The confident guess doesn't cost you anything up front. The cost shows up later, after you have made a decision based on it.

The two kinds of "not knowing"

No chatbot has documentation covering everything a customer might ask. That's just a fact of running one, not a defect in a specific product. What actually varies between tools is whether a gap in what it knows shows up where you can see it.

One kind of not knowing is visible: the chatbot says plainly that it doesn't have the answer. You know exactly where you stand. You can go ask someone, check the source yourself, or decide the question wasn't important enough to chase down. Nothing about your situation got worse, you just didn't get an answer this round.

The other kind is invisible: the chatbot doesn't have the answer either, but it gives you one anyway, dressed up in the same tone it would use for something it actually knows. You have no signal that anything is wrong. You take the answer at face value because there's nothing about how it's delivered that hints otherwise. This is the one that actually causes damage, and it's also the one that looks better in a five minute test.

A test worth running when you're comparing chatbots

1. Find a real question about the product that genuinely isn't covered in the documentation. Something specific enough that a guess would be easy to generate, but wrong.
2. Ask it to each chatbot you're evaluating, worded the same way each time.
3. Don't score the answers on whether they sound helpful. Score them on whether you could tell, just from reading the response, if the chatbot actually knew what it was talking about.
4. If it admits the gap, check what comes with the admission: a name, a link, a real next step, not just an apology. An admitted gap with a real next step is more useful than a guess you cannot identify as one.

Why this is easy to get backwards during a demo

When you're comparing a few tools side by side, the one that has an answer for everything looks more capable. It feels complete. The one that says "I don't know" more often feels like it's missing pieces the other one has. That impression is misleading, because a chatbot that never runs out of answers isn't necessarily better informed. It might just be less willing to admit the edges of what it actually knows.

The difference doesn't usually show up in a short demo. It shows up weeks later, once you've built something on top of an answer that read just as confidently as a correct one would have. By then the chatbot that seemed more complete in testing has already cost you real time, and the one that seemed incomplete has quietly been telling the truth the whole way through.

How we handle this

Our agent is built to decline rather than guess when it doesn't have a real answer in what it's grounded in, and to hand you to a real person instead of generating something that merely sounds right. We're not claiming it never misses a question. We're saying that when it doesn't have an answer, it's built to say so rather than cover the gap with something plausible.

Related reading

• See all 20 guides in the Actually Helpful library, grouped by what you are actually trying to figure out.
• The confidence problem: why chatbots rarely say "I don't know", on why this honest admission is uncommon in practice even though it's the right call.
• What should happen when a chatbot can't answer, on what a good handoff looks like once the gap is admitted.
• Why AI chatbots give confidently wrong answers, on why a wrong answer and a right one can read the same way.
• How to tell if a chatbot is reading your docs or just guessing, on the other signals worth watching for besides a plain admission of a gap.
• If your help docs are spread across five different tools, what does a chatbot actually search?, on why an honest "I don't have that" is often less about the chatbot's judgment and more about which sources it was actually given.