When the answers are free, the questions are the job

Sam Altman reckons 95% of agency work is going away. A German colleague told me that at an internal briefing, the head of comms at Audi put it at 80% within two years. Pick your figure — the direction is the same, and the evidence points at it being at least directionally true.

So here is the question I keep coming back to. If AI does almost all of the hourly services work we currently buy from people, what is the skill that's left?

The question machine

There's a documentary in which the architect Renzo Piano describes himself as a "question machine". His practice runs major projects in several countries at once, so his own job has narrowed to flying between them and checking on each. In one scene he walks the length of an office past dozens of drawings pinned up for his inspection. He stops at each one, examines it, and interrogates the person who drew it. That, he says, is his role now — not to answer his clients' requests, but to grill his team with the questions they should be answering.

If generative AI is our conduit to the whole of recorded human knowledge, then we no longer need knowledge workers as such. There's a good argument — I first saw it in a video a colleague shared — that AI is best understood as a cultural technology, in the lineage of language, writing and the library. Not intelligent in itself. A way of reaching human intelligence at scale. If that's right, the machine holds the answers. What it can't do is decide what's worth asking.

Curiosity, logic, taste

What's valuable, then, is the Renzo-like question machine. In my experience it's a blend of three things.

Curiosity, because you have to want the answer badly enough to keep asking. Logic, because you have to think sequentially about how to ask — how to structure a problem into questions, and questions into outputs. And taste, which is shorthand for having some idea of what the output should look like before you've made it. Taste requires an omnivorous intake of ideas and the judgement to curate them down.

It's perfectly possible to produce good work without taste. Just let the AI do it for you. But great work in any field — an elegantly assembled legal argument, a brand-defining campaign — needs someone who already knows what great looks like. That part doesn't come out of the model.

The assembler

In another interview, Piano describes himself as a builder, an assembler of pieces — a kind of pirate who steals elements from other buildings and fits them into a coherent whole. He compares his method to Italo Calvino, an obsessive note-taker who "stole from daily life" because he "never believed in the idea of a wild creative impulse".

That is as good a description as any of working well with AI. Someone constantly combing the world for inputs, feeding them into the prompt, and stitching a bricolage of outputs into something original and whole. The model can generate. It can't notice what's worth stealing, and it can't tell when the assembly is finished.

Which is the part I'd want a person doing anyway. The answers are becoming free. The questions — and the taste to know a good one — are the job.