We judge an AI product by whether its answers are accurate, fast, and deliverable. But there's an earlier number than any of those: how many times a day does this user want to ask something?
For most jobs that number is small. And it doesn't grow when the model gets better.
We are the biased sample
People who build AI products tend to use themselves as the user. We open a chat box dozens of times a day, because "pose a question, get an analysis, keep moving" is literally how our work runs.
Most jobs don't run that way.
The moments that genuinely call for thinking up a question, waiting for an analysis, and then deciding something might come three or four times a week for a salesperson, and less for someone in operations. The rest of the time people are moving through process, filling forms, answering mail, reconciling definitions, checking whether the other side received it — and providing their boss with emotional reassurance. All of it is real work. None of it takes the shape of a question.
It isn't that users don't know how to use AI. Their work simply doesn't generate that many questions.
It's a multiplication
Break a product's usage apart:
Usage ＝ how many people use it × how many times a week each one does
The effort we pour into better answers acts on the second term, and only on part of it: going from winning two of three occasions to winning all three. That's worth doing, and it has a ceiling — winning them all still means three.
Answer quality decides whether you take an occasion. It doesn't decide how many occasions there are. That count is set by the product's form: a product that is only an input box appears exactly when the user thinks of it, no more and no less.
Three ways to raise the count
The ceiling isn't on the model side, so the form is where you have to look.
One: get inside the task. What the user is already doing is the occasion. They're looking at a quote, an order, a shortlist — put the judgment they need right there. They don't have to ask, and don't even have to register that this is AI. They never "used" it once, and the work got done.
Two: let the system find the moment. Data changed, a date arrived, a risk appeared — none of these need the user to remember you. That's the whole subject of AI finds the scene, instead of people finding AI: taking part of the trigger out of the user's hands.
Three: make one occasion go deeper. If you can't raise the count, raise the yield per occasion. One sentence sets off a chain — fetch, compare, generate, write up, deliver — and a single question returns what used to take an afternoon. This is usually the least-eaten piece: teaching users to ask chain-shaped questions is easier than getting them to remember you more often.
Two boundaries
None of this says the chat box is useless. For people who really do have dozens of questions a day — engineers, analysts, researchers — the input box is the best possible form, and products like Claude Code are the proof. The chat box isn't the mistake; generalizing from that population to everyone is.
And don't manufacture occasions. When the count won't move, the easiest action is to push: one notification a day and the number goes up. But one wrong push costs far more than not pushing. Being proactive is earned by the moment genuinely being worth an interruption — and judging worth it is much harder than generating the content. That judgment is where the real engineering in a proactive product lives.
So when I size up an AI product, I start with a stupid question: how many times a week can it legitimately appear in front of the user?
How good the answers are comes after that.