Experience, Repriced

The verdict: right about the direction, wrong about the object. Something did get commoditized. It just was not experience.

What I claimed

Reprises usually flatter their originals, so let me restate the 2024 essay honestly. It made six claims.

1. AI commoditizes experience, making it accessible and transferable, like a product.
2. A generalist with AI can act with specialist depth; the gap between them narrows.
3. Productivity stops being about accumulated experience and starts being about how effectively you use AI.
4. Specialist tacit knowledge may not be fully replicable. (A hedge, not a claim. Hold that thought.)
5. Over-reliance could erode critical thinking, and AI bias needs human oversight.
6. Education must shift toward AI collaboration skills.

The essay also carried a hinge assumption it never discharged. The whole argument works only if the portion of experience AI can deliver is the portion that matters. I flagged that assumption in paragraph three, called it a limitation, and kept walking. Twenty-two months later, that unexamined hinge is where the essay swings.

What held

Three things aged well, and I will take them.

The framing held. "Not replacement but commoditization" located the disruption in the right place: the experience premium, not the headcount line. The first work AI absorbed was not jobs wholesale. It was the legible slice of experience inside jobs, the part that junior professionals used to sell while they built the rest. Ask anyone hiring junior analysts, junior associates, junior marketers in 2026 how that premium is doing.

The productivity claim held and then some. I wrote that success would depend less on years served and more on how well you wield the tools. That was the firmest sentence in the essay and it turned out to be an understatement. The spread between operators who can direct this leverage and operators who cannot is now wider than the spread experience alone ever produced.

And the small-team claim held. I said smaller organizations would compete on a more level field. The tiny-team pattern, a handful of people running what used to take a department, is no longer a prediction. It is a pattern you can watch.

What broke

Now the corrections, because they are more useful than the victories.

The object was wrong. Experience did not get commoditized. The evidence of experience did. Fluency, the polished artifact, the authoritative paragraph, the specialist-shaped deliverable: all of it dropped to roughly the price of electricity. Sounding like an expert became free. Being one did not. I wrote a whole essay about that gap this month, The Expert Costume, and the one-line version belongs here: AI did not make expertise easier to build. It made expertise easier to fake, including to yourself.

The word was wrong too, in a way I find funny now. In 2024 I defined commoditization as "making experience accessible and transferable," which is not commoditization. That is democratization. Commoditization, the cold market meaning, is when a thing loses its premium because supply explodes. I accidentally predicted the right phenomenon with the wrong definition: the parts of experience that can be written down lost their premium exactly the way commodities do. Which falsifies the most elegantly wrong sentence I wrote that year: "This shift doesn't diminish the value of experience." It did. It collapsed the price of every part of experience that is codifiable, and it raised the price of everything that is not.

The boundary was wrong. I scoped AI to "routine tasks." Routineness turned out to be irrelevant. The real boundary is codifiability. Plenty of work that experts swore was judgment turned out to be pattern, and the models ate it without asking whether it was routine.

And my best example argued against me. I offered the doctor using diagnostic AI as proof that generalists could borrow specialist depth. Read it again: a physician using a diagnostic copilot is a specialist amplified, not a generalist elevated. The example proves that AI amplifies whoever holds it. I had promoted an amplifier into an equalizer, and those are different machines.

What I could not see

Some of the misses were not errors. They were horizon problems, and they are the actual payload of this reprise.

Generation versus verification. The 2024 essay had no theory of what stays scarce. Here it is. Generation got cheap. Verification did not. That asymmetry is the whole story. When anyone can produce a tenth-year artifact in four seconds, the scarce input is no longer producing the artifact. It is knowing whether the artifact is right, where it stops working, and what it quietly left out.

The delay structure. I worried, vaguely, about over-reliance. The real mechanism is sharper: AI amplifies experts and exposes novices, just on a delay. The novice's costume does not tear in the demo. It tears at the edge case, in front of a client, at the worst possible moment. The 2024 essay imagined the gap narrowing. The gap narrows at the point of production and re-opens, wider, at the point of verification.

The second ditch. In 2024 I saw one failure mode: trusting the machine too much. There are two. The other ditch is older. People meet AI with the same wiring our ancestors used on the stranger at the tree line: threat, alien, the thing that comes to replace me. So they brace and wall it off. I wrote about this in AI Is the New Other, and the short version is that our own origin story says the fear is the trap, not the wisdom. Over-trust wears the costume. Under-trust refuses the lever entirely. Both lose, in opposite directions, and the 2024 essay only guarded one side of the road.

The amplifier became an operator. The strangest miss: I underestimated my own thesis. I wrote about decision support, insights, suggestions. What actually arrived is execution. Systems that do the work, not just advise on it. Whole workflows, run end to end, with a human setting direction and judging output. The 2024 essay was directionally right and an order of magnitude too timid about its own mechanism.

"Skilled," defined at last

The most load-bearing word in the 2024 essay was the one I never defined. "Skilled generalist." Skilled how? In 2024 I genuinely could not have told you. In 2026 the answer is concrete, because we have watched who actually captures this leverage. Five things:

1. Context assembly. Knowing what the system needs to know, and getting it there. The constraint moved from access to specificity: this wave penalizes vagueness in a way the previous waves never did.
2. Decomposition. Cutting ambiguous work into pieces something else can execute.
3. Delegation in writing. Describing what good looks like clearly enough that a machine, or a person, can hit it.
4. Output judgment. Seeing what came back, what is wrong with it, what is missing from it. The test I use now: can you defend this without the model in the room?
5. Accountability. Owning the result. The model is not lying to you; it has no idea and no stake. Caring whether it is true is your job, and it cannot be delegated.

Look at that list. It is not general knowledge. It is judgment, applied at the controls of leverage. Which means the skilled generalist thesis survives, but only by collapsing into something sharper: the generalist does not become a specialist. The generalist becomes a buyer of specialist-shaped output. And buying well requires the very thing I claimed was being commoditized.

The thesis, restated

So here is the 2026 version of the 2024 claim, with the hedges converted to verdicts.

In 2024 I said the boundary between novice and expert would blur. At the surface, it did: everyone's output now reads as stage five. Underneath, the boundary sharpened, because fluency used to be the test and now fluency is the noise floor. The signal moved to the only place it could: what you do when the answer is not written down yet.

The lever is cheap. The judgment is the moat. In 2024 I told you experience was becoming a product. The correction: the product was the costume. Experience was never the inventory. It is the buyer's eye.

The audit

I will close the way the original should have: with something falsifiable enough to grade in 2028.

If I am right this time, the next two years widen the spread between firms that protected judgment and firms that outsourced it. The discount is real and everyone gets it. The moat is what you do with the freed hours.

So, two lists. First: where in your business are you still paying the experience premium for output that is now effectively free? Take the discount, today. Second: where are the freed hours going? If they are not going into the reps that build judgment, yours and your team's, then you took the discount and handed back the moat.

The words got cheap. The judgment did not. Spend accordingly.

Original essay: "AI Empowers the Skilled Generalist with Commoditized Experience," August 22, 2024. Companion pieces: Leverage You Operate With Words (April 2026), AI Is the New Other (June 2026), and The Expert Costume (June 2026).