The data suggests something the headline did not: the number arrived without a date. An article circulated recently claiming that Anthropic staff expect AI to handle 75% of their work by 2030 landed in my feed via Crypto Briefing, an outlet whose editorial corpus is otherwise devoted to token prices and protocol upgrades. The piece contained no crypto content whatsoever. That absence is the first tell. What we have is not a research finding. It is an internal sentiment poll, laundered through two media hops, republished by a vertical that needed AI traffic and got it.
Let me trace the chain. Anthropic employees, presumably surveyed or quoted internally, express a belief about 2030. A technology outlet reports it. Crypto Briefing rewrites the report. By the third hop, methodology is gone. Sample size is gone. The phrasing of the question is gone โ and this matters enormously, because 'tasks' and 'hours' and 'headcount' are not the same denominator. The publication date is also gone, which is the most damaging omission of all, since it prevents any alignment against the capability curve at the moment the claim was made.
None of this makes the number false. It makes it unlocatable. And an unlocatable number that supports a $300 billion valuation is not a neutral artifact.
The mechanics of task-duration forecasting deserve scrutiny before we accept any 2030 projection. METR's work on task-length horizons found that the duration of tasks models can complete at 50% reliability has been doubling roughly every seven months, with a baseline near fifty minutes of human work at the time of measurement. Extrapolate that linearly from 2025 to 2030 and you land at eight or nine doublings โ hundreds of hours, meaning multi-week human projects. On that arithmetic, a claim that AI handles three-quarters of knowledge work is not fantasy. It is a defensible extrapolation.
But extrapolation has a known failure mode, and I have watched it fail before. In 2020, I spent six weeks reverse-engineering MakerDAO's collateralized debt position mechanics on a local Ganache node, stress-testing liquidation cascades under volatile ETH prices. The math was clean. The oracle latency edge case I found was real. What the model could not capture was the behavioral response of liquidators once the incentive structure shifted mid-cascade. Linear models of nonlinear systems produce confident numbers right up until the moment the system changes shape.
The same applies here. Three fractures are already visible. First, the data wall and diminishing marginal returns on compute. Second, the reliability gap: going from 50% to 99% correctness โ the actual enterprise threshold โ is not a linear continuation, it is a different engineering problem with different economics. Third, error compounding along long task chains: a 99% per-step success rate across a hundred steps yields roughly 37% end-to-end success. That is not a rounding error. That is the whole ballgame.
The strongest counter-evidence is empirical, not theoretical. METR's July 2025 randomized controlled trial found experienced open-source developers using AI tooling took about 19% longer to complete tasks while subjectively believing they were 20% faster. Read that sentence twice. The perception of acceleration and the measurement of acceleration moved in opposite directions. And '75%' came from a perception survey. It is drawn from the same well as the 20%.
There is also an Amdahl's law trap sitting inside the framing. Automate 75% of subtasks and the remaining 25% โ judgment, accountability, cross-party coordination, exception handling โ does not compress proportionally. If that residual becomes the bottleneck, total hours saved falls far below 75%. Task automation rate is not equivalent to hours-replacement rate, and neither is equivalent to headcount reduction. The article collapses all three into one decimal.
Behind the collateral lies a maze of incentives, and here they point in one direction. Anthropic derives revenue from enterprises paying to substitute or augment labor. An employee base predicting that the company's own product will absorb 75% of work is functionally a supplier forecasting 75% market penetration for its own good. In serious equity research, that disclosure would carry a methodology appendix and a heavy discount. Here it carries neither.
The commercial function is legible if you follow the pricing. The company's revenue center of gravity has shifted from raw API access toward coding and agentic workflow products, where the pricing logic is anchored to displaced labor value rather than token volume. If a procurement officer believes AI handles 75% of the work, the ceiling on willingness to pay moves from the tooling budget to the payroll budget. That is roughly a tenfold expansion in addressable price. The number is not a forecast. It is pricing infrastructure.
And it is necessary infrastructure, because agentic workloads consume tokens at multiples of chat workloads โ long context, multi-turn tool invocation, retry on failure. Priced against displaced labor, gross margin holds at software levels. Priced in a race with OpenAI and Google, margin collapses while usage explodes. The narrative exists to prevent the second outcome.
Notice also the delivery mechanism. The signal was released as employee expectation rather than executive projection. That is a designed deniability: challenged, it degrades gracefully into 'individual staff sentiment from an internal survey.' Executive forecasts do not have that escape hatch. The choice of mouthpiece is itself a strategic decision.
I do not trust the doc; I trust the trace. And the trace runs through a media vertical with no domain expertise in labor economics, incentivized by click-through, amplifying a figure from a party with a direct financial interest, with no adversarial check anywhere in the loop. Two parties, aligned interests, zero verification.
The labor market will not wait for methodological rigor. Stanford's Digital Economy Lab found in 2025 that employment among 22-to-25-year-olds in AI-exposed occupations fell roughly 13% relative to trend, while senior workers were essentially unaffected. That asymmetry โ junior collapse, senior augmentation โ is the real shape of the transition, and it is already underway. It does not require 2030. It does not require 75%.
The residual 25% is where the argument actually lives. If the remaining quarter of the work consumes 60% of the hours, then organizational compression is a fraction of the headline. Nobody selling agentic tooling wants to publish that arithmetic, and the survey question almost certainly did not ask it.
The infrastructure constraint is the one nobody calculated. Shifting 75% of knowledge work to inference-heavy agentic load implies a token profile an order of magnitude beyond chat, and therefore power and memory-bandwidth demand that may exceed available supply by 2030. Current capital expenditure flows toward training-grade clusters โ high interconnect, high precision โ while agentic inference rewards memory capacity, cache hit rates, and scheduling density. These are not the same machine. If the physical supply does not arrive, the forecast is vetoed by physics rather than by capability, and no press release survives a substation.
So here is my read, split by proposition. The probability that Anthropic employees will still report a figure at or above 75% in 2030 is high โ narrative lock-in and group coherence are self-sustaining. The probability that measured AI work-hour share actually reaches 75% is perhaps 20 to 30%, supported by trend extrapolation but discounted by reliability thresholds, verification costs, organizational inertia, and compute ceilings. The probability that headcount contracts by 75% on that basis is under 5%. The nonlinear decay between task automation and staffing is enormous, and it is where every naive projection dies.
What should be tracked instead is far more prosaic. The ratio of directive to iterative conversation in the Anthropic Economic Index. Merge rates and rework rates on AI-generated code in production repositories. Net revenue retention on enterprise agent products โ the only number that tests whether substitution pricing survives contact with a renewal cycle. Junior employment data in exposed occupations, quarterly. And the next METR horizon update: if the doubling period lengthens materially, the 2030 case weakens in the only currency that matters.
The real risk is not that 75% is wrong. It is that a number with no sample, no wording, no date, and no dissent gets used to justify hiring freezes, curriculum decisions, and policy allocation that cannot be reversed in a four-year feedback loop. Capability expectations have become a tradable asset. Labs mint them, media distribute them at zero cost, capital prices them, and workers absorb the adjustment. No governance framework currently covers the disclosure ethics of capability forecasting, only model outputs. That gap is the finding.
When abstraction fails, the workers bleed value. The question worth asking is not whether machines will do three-quarters of our work by 2030. It is whether anyone will be held to the number when the answer arrives.