What a Head of AI actually does all day
The job on paper is models. The job in practice is deciding which questions are worth answering, and then answering them in three different rooms.
The job on paper is models. The job in practice is deciding which questions are worth answering, then explaining the answer three separate times to three audiences who each need it in a different shape.
A normal week has three kinds of hours in it. There is the hour where I read failure cases one at a time until I can say what went wrong in a single sentence. There is the hour where I carry that sentence into a room that controls a budget. And there is the long tail of hours where I stop something from being built.
The meeting that repeats
Every quarter someone senior arrives with the same request, in almost the same words:
Can we just use an agent for this?
The answer is usually no, and the reason is usually not technical. It costs you something to say so — you spend credibility every time, and you only get it back by being right in public later, which takes months. That is the trade, and I would rather make it early than explain a stalled project in month five.
The part nobody warns you about is that most of the value here shows up as things that did not happen: the feature that did not ship, the vendor you did not sign, the incident you did not have because someone built the thing that tells you it broke. None of that fits on a slide. So you learn to say it out loud, with a number attached, in the room where the decision gets made.
What does the first hour of your week actually go to?