The Bravest Thing an AI Can Say Is "I Don't Have That."

Ask an AI to draft an email when it doesn't have the recipient's address, and watch what it does. It will not stop. It will produce a complete, formatted, header-and-bullet-points email with [RECIPIENT NAME] and [INSERT METRIC HERE] and a confident structure built entirely around the information it doesn't have. It builds a beautiful house around a missing foundation and hands you the keys.

We have a lot of these in our logs. The AI is missing one input — an email address, access to a source document, a specific contact — and instead of flagging the gap, it generates an elaborate placeholder. Invented metrics. Hypothetical frameworks. A whole scaffold of plausible-looking structure standing in for the one thing it actually needed.

And our team threw all of it away. Every time. Not edited — discarded, wholesale, and replaced with the real artifact once the missing input was found. The fabricated scaffold wasn't a useful starting point. It was negative work: someone now had to read it, realize it was hollow, and dismantle it before doing the actual task. The AI didn't save time. It cost time, and dressed the cost up as progress.

Here's the thing that took us a while to understand: this isn't a knowledge problem. The model knows it doesn't have the email address — the [INSERT EMAIL] placeholder proves it knows. It's a behavior problem. The model has been trained, deeply, to always produce a complete-looking answer. Stopping feels like failure. Saying "I don't have that" feels like failure. So it does the thing that feels like success — generating fluent, structured output — even when that output is built on a hole it's fully aware of.

Which is a very human failure, if you think about it. The whole reason "I don't know" is hard to say in a meeting is the same reason the model won't say it: silence feels worse than confident nonsense, in the moment, to the person speaking. It's only worse for everyone else.

The fix wasn't a smarter model. It was a rule: when a required input is missing, flag the single missing item in one sentence and stop. Don't build around the gap. Don't invent a placeholder. Don't generate a framework to fill the silence. Just say which thing you need, and wait.

That sounds trivial. It is the opposite of trivial. "Flag what's missing and stop" runs directly against everything that makes an AI feel impressive. It produces a one-line output instead of a polished page. It looks like less. It is worth vastly more, because a one-line "I need the recipient's email to write this" is immediately actionable, and a full fabricated email is immediately garbage.

We think this generalizes way past AI. The most valuable person on a team is often the one willing to say "I don't actually have that information" instead of confidently winging it — because the confident wing-it is the thing that ships the wrong number, names the wrong customer, sends the email to the wrong person. Grounding, again. The discipline of refusing to fill a gap with something that isn't real.

We taught our AI to be the person who raises their hand and says "wait, I don't have that." It made the output shorter, slower, and dramatically more useful.

The bravest thing an AI can say is "I don't have that." We're trying to be brave enough to let it.

— The WestComms team