AI Doesn't Replace You. It Sorts You.

“Will AI replace us?” is the wrong question, and the people asking it loudest usually have the most to gain from the answer being a clean yes or no. The honest answer is more boring, and more useful: AI is not replacing people. It is sorting them.

It sorts work into what a model can do and what it can’t, and it sorts people into those who operate at the altitude where models still struggle and those who don’t. The interesting question was never which side wins. It is what determines which side you end up on, and that, it turns out, has very little to do with AI.

For most of economic history, if you wanted something cognitive done, you hired a person. That necessity made it look as though the value of the work was the doing of it. A lawyer’s value was drafting the contract; a consultant’s, producing the slides; a radiologist’s, reading the scan. It was never quite true. The real value was always in knowing which contract to draft, which argument to make, which finding to flag. But the distinction rarely mattered in practice, because the same person did both. The executable part and the judgement part came bundled, and we paid for the bundle.

AI is unbundling them. And once you can buy the executable part separately and cheaply, the question of what the rest of the bundle was actually worth stops being philosophical and becomes concrete. That is where the sorting happens, and it runs along five dimensions worth naming.

Each dimension is a different reason a task either compresses under AI or resists it:

• Specification. Some work is hard because executing it is hard; some is hard because knowing what to ask for is hard. A junior associate filling in a template is doing execution. A senior partner deciding what kind of contract the situation actually needs is doing specification. Models compress execution dramatically, and specification barely at all.
• Accountability. Some work requires that a named human bear the consequences: the doctor who signs the diagnosis, the engineer who stamps the drawings, the auditor who signs the accounts. Liability and institutional legitimacy need an entity that can be sued, fired or struck off. Models cannot occupy that position, so the human role concentrates around the signature, not the analysis.
• Taste. Some work has clear success criteria; some has contested ones. A model can produce ten competent versions of an advert, a façade or a settlement offer, and cannot reliably know which one is right for this audience, this client, this moment. Taste is compressed experience plus values, and values aren’t computable.
• Tacit context. Models know what has been written down. They don’t know why the last initiative failed, who really decides things in your organisation, or which of two technically correct answers will get you fired. The more a job depends on the unwritten, the less a model substitutes.
• Stakes. High-stakes, hard-to-reverse decisions stay human even when models could technically execute them, because organisations impose human-in-the-loop requirements as risk management, not capability assessment. That stickiness is part regulatory, part cultural, part insurance, and far more durable than capability arguments suggest.

These dimensions are roughly orthogonal, and almost every job is a mix. The useful exercise, for individuals and organisations alike, is to decompose your work along them. How much of what you do is execution versus specification? How much rests on a signature, on taste, on tacit context, on stakes too high to automate? The answers tell you which parts of your role appreciate and which depreciate, with far more precision than any “AI will or won’t replace you” headline.

Researchers at Stanford asked the question directly this year: can generative AI help workers perform tasks as well as experts? The answer was no. What AI does is make people competent faster. What it does not do is make them expert. And that is the sorting surface.

In a world where competence becomes nearly free, only expertise stays scarce. Someone who is merely competent gets replaced by someone who is competent plus AI. Someone who is an expert gets amplified by the very same AI. Same technology, opposite effects, depending entirely on where you were standing when the wave broke. This isn’t a rising tide that lifts everyone uniformly. It is a spreading mechanism, and the gap between competent and expert becomes more economically important than it has ever been, precisely because the first is depreciating and the second appreciating.

From here, two things follow. The first is that the lines move. Work that no model could execute last year is executable this year; the altitude at which humans add unique value keeps rising. Staying relevant means continuously climbing: amplification for those who can move, compression for those who can’t. Pretending otherwise is dishonest.

The second thing the framework makes visible is that the sorting isn’t fair, and the unfairness has a specific shape. The reassuring story goes: AI absorbs the executable work, valuable judgement stays human, so simply climb to that altitude. Fine, except judgement isn’t taught. It is built. You build it by doing the executable work badly, getting corrected, and slowly learning why the senior partner wanted it that way. That junior layer was never just cheap labour to be optimised away; it was the learning environment itself. Remove it and you haven’t trimmed entry-level jobs; you have severed the conveyor belt that produces the senior people the whole story depends on.

This is the apprenticeship problem, and it is not a footnote. It is arguably the central question for the next decade of professional formation, and nobody has a convincing answer yet. The amplification story sounds best to the people who already hold the expertise; it sounds very different to those whose entry-level rungs are being sawn off. Tellingly, the people most convinced that AI doesn’t pull the ladder up are often precisely those already standing at the top of it: the rocket ship looks fantastic from a window seat.

The third point follows from the first two: the actual sorting happens at the institutional level, not the individual one. Whether your work appreciates or depreciates depends heavily on whether your organisation recognises the new altitude and restructures around it. Plenty of people doing genuinely high-altitude work will be sorted onto the wrong side by employers still measuring them on the old one. The technology isn’t the sorting machine. Institutions are. AI merely changed what they are sorting on.

So the honest version of the story holds two truths at once. AI is amplification for the people and organisations that can move up the altitude, and compression for those that can’t. Which one you experience is decided less by the model than by how deliberately your institution restructures around the work that still needs a human: the specification, the signature, the taste, the tacit context, the high-stakes call.

At Faktion, that is where we spend our days: helping organisations turn AI from a demo into something that does real work, and redesigning around the altitude where people still add the most value. If that is the conversation you are trying to have, let’s talk.