Where I let the AI decide, and where I don't

The question I get in interviews is whether I use AI. The useful question is a different one: where do I let it decide.

I split the work on one criterion. The agent is fast at anything whose right answer is written down somewhere, and dangerous at anything that needs taste. So it counts, converts, propagates and checks. I decide the measurement, the hierarchy and the word.

Counting looks like the harmless part and it is where I have been wrong most. I announced a component count based on layer names, and the design file numbers duplicates: forty became seven once I started counting by the component's identifier. In another census I included hidden layers and inflated the total by 28%. Both times the agent did the right arithmetic on the wrong question, which is the only thing it knows how to do.

Propagating is where the damage is fast. A script of mine swept an entire repository to fix one color pattern, fixed it, and destroyed another pattern it did not know existed. A global sweep has no notion of what it does not know. Today every bulk change arrives with the list of what changed, and I read the list before I accept it.

What I do not delegate. What the measurement of the problem is, what is in scope and what is out, which word goes on the button, and which of the two screens is right. This is not craft pride. It is that the agent has no way of knowing the ruler is wrong, because I am the one who handed it the ruler.

The real gain also showed up somewhere I did not expect. It was not drawing faster. It was being able to ask the repository what used to cost an afternoon: which screens use this measurement, how many are still outside the language, what breaks if I change this value. The 39 combinations of one component I delivered on Directo's feed came out of that, with no new measurement in the system, because "how many measurements do I already have" became a question with an answer in seconds.

AI did not give me taste. It gave me reach.