When Execution Becomes Almost Free: A New Definition of Talent

By PeterZou

I. Two Paradoxes That Arrive at the Same Time

Start with two sets of data that, placed side by side, make for uncomfortable reading.

PwC's 2026 Global AI Jobs Barometer shows that the wage premium for AI skills has climbed to 62% (up from just 25% in 2024 and 56% in 2025) — "knowing how to use AI" really does pay. 【verified】Yet the same report also shows that in 2025, Belgium's "AI user-type" roles fell 5.7% and "AI developer-type" roles fell 26.4%, while global hiring for AI specialists grew at roughly 8 times the rate of overall hiring. 【verified】

This looks self-contradictory: if AI skills are so valuable, why are the related jobs shrinking?

The answer is that what is valuable is no longer the skill itself, but the scarcer thing behind it. And the way most of us think about talent is still stuck on the old foundation of "skill stock." "Knowing how to use a given tool" is both a ticket of admission and a fast-depreciating asset; what actually endures is cross-tool judgment, not proficiency with any single tool. 【inference】

This essay is about precisely that: when the marginal cost of execution approaches zero, the anchor of talent value is migrating from "stock of skills" to "goal definition + tool orchestration + verification and accountability."

II. The Criterion: What Actually Counts as a Revolution in "Talent"

We first have to establish a yardstick, or everything will look like mere deepening rather than revolution.

A paradigm revolution in Kuhn's sense is not about doing the old things better; it is about changing the standards for what counts as a problem and what counts as a good answer. Applied to "talent," we need ask only two questions: Has the unit of account for talent value changed? Has the warrant changed?

If the change is merely "using AI to do the same skill faster, to perform the same job better," then the unit of account is still the stock of skills a person possesses and their output per unit of time, and the warrant is still the certification of degrees, experience, and seniority — that is only deepening. Only when the unit of value migrates from "what skills I have and how much I produce per unit time" to "how reliably I can convert a goal into a verifiable result, and be accountable for it"; and when the warrant migrates from "credentialing" to "the result is verifiable + responsibility is assignable" — only then is it a revolution. 【inference】

With that yardstick in hand, let's look at the evidence.

III. The Evidence: The Old Assumptions Are Loosening on Five Fronts at Once

The old view of talent rested on five self-evident assumptions: talent is a stock of skills (Becker's human capital theory); it is packaged into job task bundles (from Smith's division of labor to modern HR); it appreciates through accumulated experience (Arrow 1962, "learning by doing"); it is certified by degrees and credentials (Spence's signaling theory); and it is owned by the organization (McKinsey's 1997 "war for talent"). 【verified · classic literature】

AI is shaking each of them, one by one.

First, execution skills are becoming a public good. In a randomized controlled trial of GitHub Copilot, the treatment group completed tasks 55.8% faster than the control group, and the less programming experience people had, the more they benefited. 【verified】In a pre-registered experiment in Science, Noy and Zhang had 453 college-educated professionals perform professional writing tasks; the group using ChatGPT cut time spent by 40% and raised quality by 18%, with weaker performers gaining the most. 【verified】Brynjolfsson et al.'s field study of 5,172 customer service agents is even more direct: an AI assistant raised issues resolved per hour by +15% overall, +36% for the lowest skill quintile, while the highest-skilled saw almost no gain — and a slight decline in conversation quality; new employees reached in two months the proficiency that non-AI users reached in six. 【verified】

Second, experience does not equal leverage. METR's randomized controlled trial recruited 16 seasoned open-source developers and had them complete 246 real issues: when allowed to use AI, they actually spent 19% more time; developers had forecast a 24% speedup beforehand and still believed they had been 20% faster afterward — perception and reality pointed in opposite directions. 【verified】Experience must be recalibrated before it can be converted into AI leverage; otherwise proficiency becomes a source of overtrust.

Third, the real core capability becomes "knowing where you will be wrong." Dell'Acqua et al.'s pre-registered experiment at Boston Consulting Group involved 758 consultants: on tasks within AI's capability, those using GPT-4 completed 12.2% more tasks, were 25.1% faster, and produced quality more than 40% higher, with below-average performers improving 43% versus 17% for above-average performers; but on one task deliberately chosen outside AI's capability boundary, consultants using AI were 19 percentage points less likely to answer correctly. 【verified】This is the so-called "jagged frontier" — only by using AI extensively do you learn which side is which. The new core of talent value is the judgment to place a task on the right side of the frontier. 【inference】

Fourth, verification and accountability become the new bottleneck. When generation becomes cheap, the constraint shifts to proving that it is right. In the customer service study above, the strongest performers actually saw quality decline after adopting AI suggestions — however good the suggestion, it cannot replace one's own judgment and responsibility. 【verified】Scarcity has not disappeared; it has moved from "access and execution" to "verification and accountability."

Fifth, the pipeline that reproduces talent is breaking. Stanford Digital Economy Lab, using high-frequency ADP payroll data, found that employment levels for 22–25-year-olds in AI-exposed occupations are 19% below counterfactual, while experienced workers show no comparable gap, and the difference operates mainly through reduced hiring rather than layoffs. 【verified】The Atlanta Fed connects this back to Arrow: entry-level jobs are "not just low-value work, but the curriculum through which workers accumulate the human capital they need for future productivity"; by automating them away, firms are sawing through their own future supply of senior talent. 【verified】This is the most dangerous and the least discussed point — not that experience is worthless, but that the path to acquiring it is being closed.

Meanwhile, the scarce resource has moved upstream. PwC 2026 found that AI-exposed entry-level roles are 7 times more likely than low-exposure roles to require traditional senior skills such as leadership and strategic thinking, and 2.5 times more likely to add tasks that depend on "human-intensive skills" such as empathy, creativity, and judgment. 【verified】Even employers' hiring standards are shifting — less emphasis on formal degrees, and more on adaptability, technical fluency, and problem-solving. 【verified】

IV. The Deeper Impact: A New Definition of Talent

Gathering these threads, the new paradigm's operational definition reduces to a single multiplicative formula:

Talent value = goal-definition capability × tool/agent orchestration capability × verification capability × accountability capability, applied to the right task.

It multiplies; it does not add. Being able to define a goal but not orchestrate tools is idle dreaming; being able to orchestrate tools but not verify is a hallucination amplifier; being able to verify but not take responsibility means the value cannot be priced by the market. If any one factor is zero, the whole collapses. 【inference】

Three implications follow:

The asset has changed. The asset an individual can carry is no longer "the stock of skills in their head," but a combination of judgment and taste + private verification corpora/eval sets/workflows + orchestration protocols. Skills can be replicated by models, but these three are bound to a specific domain, specific clients, and a specific history of failure — and are the hardest to make public. 【inference】

Pricing has changed. Talent moves from being priced "by skill stock / hours" to being priced "by verifiable results and leverage." An individual's marginal value = the conversion rate at which AI capability becomes verifiable business results, not personal output speed. 【inference】

Organizational form has changed. Talent shifts from "filler for a job" to "definer of goals + orchestrator of human-machine systems + bearer of responsibility"; the employment relationship moves from "buying time / skills" to "buying results / judgment," and the market form of talent spreads from employment toward contracting and entrepreneurship. Stripe has observed and publicly argued that "independent entrepreneurship is at an all-time high." 【verified · vendor claim; specific magnitude to be verified】This is entirely isomorphic with the one-person company (OPC): an individual's effective span of work expands from "how much I can execute" to "how much I can define and verify."

V. What to Do

For individuals: stop hoarding skills that expire; start accumulating assets that cannot be made public. Let every piece of work settle into a private eval set, a failure-case library, and a record of judgment; deliberately practice the metacognition of "placing tasks on the right side of the jagged frontier"; and actively track where you systematically go wrong, rather than only remembering how powerful AI is.

For organizations: redesign your "entry-level curriculum." If entry-level roles really are disappearing, you must supply the lesson of "learning by doing" some other way — reinvesting the time AI saves into judgment and verification training for newcomers, rather than simply cutting the pipeline. At the same time, performance and pay should gradually shift from "hours / skill matrix" to "verifiable results + quality of judgment."

For researchers and founders: treat the "verification tax" as an opportunity. Whoever can provide credible verification, evaluation, and accountability mechanisms for a given human-machine collaboration scenario holds the bottleneck of the new paradigm. Goal-definition tools, orchestration protocols, private verification corpora — these are all positions not yet fully occupied.

VI. Conclusion

This revolution is not yet complete, but it has already passed the threshold at which the definition must be rewritten.

The real watershed is not "will AI replace people" — that is a question of capability; it is "when the marginal cost of execution approaches zero, will the anchor of human value migrate from execution capability to definition, judgment, verification, and accountability" — that is a question of criterion.

Skills have not disappeared. They have gone from "a scarce stock that is purchased" to "public capital that can be called upon at any time." What is scarce has, for the first time, come to rest this clearly on goals, judgment, verification, and responsibility.

The new definition of talent comes down to one sentence: it is not what you know, but what results you can be responsible for.