It is a Monday morning, two days before I give an AI lunch-and-learn to a product company with a few dozen people. My contact there sends me a message. Did I get the email?
Their team sent me the prep package on Friday afternoon: a spreadsheet, a video walkthrough, and the 16 questions they want me to answer. I search my email for it.
Nothing.
I ask them to send it again. Then I find out where the first one went.
It never had a chance.
The AI filter I built to protect my attention had flagged it as spam.
My AI had decided the prep material for an AI talk was cold outreach.
I know why it happened. I get a lot of cold email, and the pitches that get through are the personal ones. “Nhan, I saw you went to Georgia Tech, let’s connect.” AI does the research and writes the flattery, so an email that sounds like it knows me is exactly what spam looks like. A real client writing to me for the first time sounds the same.
That part I can forgive. A person could make the same mistake.
What I keep coming back to is what it didn’t do. It didn’t set the email aside as a maybe. It didn’t leave it in the inbox with a note. It didn’t say, “I’m not sure about this one.” It made the call instantly, with no doubt at all, and told no one.
I went into that week thinking the risk with AI was a wrong answer. I came out of it thinking the risk is a wrong answer that sounds exactly like a right one.
My contact told me the story had to go in the presentation. It did, as the answer to question 15: what is AI bad at?
The honest list
Here is what I told that room.
It is confidently wrong, and it never tells you. The spam filter is the whole lesson. A wrong answer from AI arrives in the same tone, at the same speed, with the same certainty as a right one. There is no hesitation in its voice to warn you.
It doesn’t know what it doesn’t know. AI works from a context window, a fixed amount of space that holds everything it can see for the task in front of it. Whatever is in the window is all it knows. Leave something out and, as far as the AI is concerned, it doesn’t exist. Put in too much and it starts to forget. Either way, it will confidently treat what is in that window as reality, even when reality is somewhere else. It cannot know a vendor is about to miss a ship date because their plant is closed for a holiday. That lives in a phone call.
It can’t own a decision. It can recommend the order. It can’t sign the purchase order, and it can’t answer for it when the wrong container shows up. Somebody’s name goes on the number.
Point it at a broken process and you get a faster broken process. AI does what the process does, only more of it and quicker. If the process is wrong, fix it first. Then hand it over.
Three of those four are the same problem in different clothes. The machine does not raise its hand. It doesn’t say it is unsure, it doesn’t say it is missing something, and it doesn’t say the process it was handed is broken.
Plan for the silence
You can’t make AI tell you when it is unsure. You can build your work around the fact that it won’t.
Check where it puts things, not just what it says. My filter’s decisions were invisible until a person asked about a missing email. Anything that sorts, files, or deletes on your behalf needs a place you look on purpose.
Give it what it is missing. Before you ask, ask yourself what it can’t see: the phone call, the vendor’s history, the thing everyone in the building knows and nobody wrote down. Put it in the window, or keep that part of the decision for yourself.
When it is wrong, teach it in plain words. The fix for my filter isn’t code. It is a sentence: a client writing to me for the first time is not cold outreach.
The filter will learn that. It still won’t tell me the next time it isn’t sure.
AI will never say “I’m not sure.” Keep someone in the room who will.
