Ask someone what makes an expert and they will usually describe a stockpile: years of accumulated facts, cases, and patterns held in one person’s head. That mental model has organised how we hire, credential, and pay for knowledge for a century. It is now out of date.
A capable model has read more filings, papers, and transcripts than any individual will touch in a career, and it will recall them on command. The stockpile is no longer scarce. So the question is not whether expertise survives — it is what it becomes once recall is free.
The half we could bill for
Recall was always the measurable half of expertise — the half you could examine, certify, and put on an invoice. The other half never fit on a résumé: framing the real question, smelling when the tidy answer is about to be wrong, deciding under uncertainty when the evidence runs out. We leaned on the billable half because it was legible, not because it was the valuable part.
AI is brilliant at the legible half and structurally weak at the rest. That is not a temporary gap to be closed by the next model — it is the line along which the work is being re-sorted.
Skills that appreciate, skills that depreciate
If you want a rough hedge against the shift, watch which muscles a model can copy and which it cannot. The depreciating skills are the ones built on retrieval and synthesis of what is already written. The appreciating ones are harder to name because we rarely trained for them directly:
- Question framing — turning a vague worry into the one question whose answer actually changes the decision.
- Verification instinct — knowing which claims need a primary source and which are quietly load-bearing.
- Taste under uncertainty — judgment about what to do when the data is thin, conflicting, or absent.
- Ownership — being accountable for a call, which a model can inform but never carry.
The expert of the next decade is not the person who knows the most. It is the person who knows which question to ask, which answer to trust, and when to stake a decision on it.
From holder to orchestrator
The practical shape of this is an expert who works more like a conductor than a soloist. The model handles breadth — drafting the landscape, surfacing the sources, flagging what it cannot verify. The human supplies the parts that only lived experience contains, and stitches the whole into a decision someone is willing to own. Counter-intuitively, that makes generalists valuable again: when the specialised facts are a query away, the scarce skill is connecting fields and holding the whole picture.
None of this erases deep expertise. It concentrates it — onto the questions and judgments machines cannot reach, and away from the recall they now do for free. The experts who feel threatened are usually defending the half that was never the point. The ones who feel liberated have already moved up the stack.