The Nudge That Was Wrong 26 Times in a Row

For twenty straight days, my own automation told me to go sign a document. Twenty-six separate nudges, each one a little more insistent than the last. Sign it now. This is still open. You need to act on this.

The document had been signed on day one.

I built the thing that was nagging me, so I got to watch exactly how it fooled itself. My verifier checked my email for a confirmation that said “signed.” Reasonable, except the portal that actually held the signature never sends one. No email ever arrived, because no email was ever going to arrive. The system read that silence as “still open” and kept escalating. Absence of proof quietly became proof of absence, and it kept demanding I act on something already done.

This is the failure mode of every eager assistant. Not that it is wrong. That it is confidently wrong, and it dresses up a gap in its own knowledge as a fact about the world. “I found no confirmation” got translated, somewhere in the machinery, into “you have not done this.” Those are not the same sentence. One is about the document. The other is about the assistant’s own blind spot.

So I changed the rule. Before it tells me anything is outstanding, it now has to load the real page. The actual source of truth, not a proxy for it. Not my inbox, not a cached status, not the absence of a notification. The signature lives on a portal, so it reads the portal. If it cannot reach the portal, it does not get to guess. It says “I could not verify this,” and it hands the decision back to me.

That last part matters more than the fix itself. An assistant that will admit “I do not know” is doing something harder than it looks. The easy move is to fill the silence with a confident answer, because confidence reads as competence. But a system that never says “I could not check” is a system that will eventually nudge you twenty-six times for nothing, and then you stop trusting the one nudge that is real.

I would rather it interrupt me less and mean it more. This is the same contract I want from any agent that acts for me: show me the source, or tell me you could not find it. Never hand me a conclusion built on a blank.

An AI that will not guess is more useful than one that is confident and wrong. Twenty-six times over, I have the receipts.