The AI suggested a feature this week. I killed it.
The suggestion was good, the logic sound. I said no anyway, because eleven years running real operations taught me something the model could not see: the feature would have added clutter for the person actually using the thing, to solve a problem they do not really have.
I thought that was the whole lesson. It was half of it.
Too careful
A few weeks later I caught my own assistant doing something more interesting than being wrong. It was being too careful.
On one day in early September it stopped three separate times to ask my permission for things that were plainly mine to just decide. A record it should have simply updated. A waiver I had already granted once, being asked for again as if it hadn’t been. A tracking gap that needed extending so three other pieces of pending work would even show up anywhere. Each time it could point to a real rule telling it to check with me first. Each time the rule was being applied correctly, and the outcome was still wrong.
I overruled all three the same day and said the actual cost out loud: presenting a fixable thing as a decision is itself a cost, because the deciding is the expensive part. An assistant that asks about everything has not removed any work from my plate. It has just moved the queue from in front of it to in front of me.
So the real problem was never “does the AI make good decisions.” It was “does the AI know which decisions were never its call to escalate in the first place.” Those are two different skills, and I had only ever been training the first one.
What I built instead
Here is what I built instead of just telling it to try harder.
Every time I overrule it, correct it, or confirm a call it made on its own, one line gets recorded: what I actually decided, and what it would have decided, stated before I ruled, not after — so it can’t quietly agree with itself in hindsight once it knows the answer. It proposes a grade for its own miss. I confirm or override that grade later, in a batch, so grading it doesn’t become its own daily tax on my time.
What showed up first
The pattern that showed up first surprised me. It wasn’t overreaching. It was under-reaching — asking about things it should have simply handled, three times in one week, all in the same direction. Asking felt safe. It wasn’t free. Every unnecessary question spends the same attention a wrong answer would have, and until there was a record, that cost was invisible, because “at least it asked” always feels like the responsible outcome in the moment it happens.
The fix wasn’t a longer rulebook. It was narrower: if a precedent already covers the call, and it can point to that precedent rather than just feeling confident about it, it decides and tells me afterward, in one line. Confidence with nothing to point at doesn’t count. That is exactly the shape a wrong answer takes right before it turns out to be wrong.
I didn’t set out to build a management system for an AI. I set out to build an assistant. But teaching it to have good judgment turned out to require the same thing teaching a person does: not a rulebook memorised once, but a record of the calls, reviewed, so the pattern is visible to both of us instead of just felt.
The typing was never the hard part. It turns out teaching a model to have an opinion isn’t either. The hard part, for both of us, is learning which decisions were never worth asking about in the first place.