A number entered the discourse last quarter and refused to leave: Anthropic staff expect AI to handle 75% of their work by 2030. It circulated through Crypto Briefing, then aggregators, then LinkedIn posts, then board decks. I have read the original item four times. It contains three sentences of substance. No sample size. No methodology. No publication date. No questionnaire wording. No distribution.
The number is real in the sense that someone said it. It is not real in any sense that survives audit.
So let me do what the aggregation chain did not: treat the number as a data point about the people who produced it rather than about the future of labor. That distinction is the whole article.
The forecast is a commercial asset, not a research finding. This is not cynicism — it is arithmetic. Anthropic's revenue comes from enterprises paying to substitute or augment human labor. An employee of that company predicting that the product will absorb three quarters of all work is a supplier forecasting a 75% market penetration rate for the supplier's own good. In serious equity research, that sentence would require a methodology appendix and a discount. In crypto media, it ran as a headline.
Here is the mechanism nobody priced. Agent-class products — Claude Code, computer-use tooling, long-horizon task runners — are priced on substituted labor value, not on token cost. If a customer believes AI absorbs 75% of a role, their willingness to pay shifts from the tooling budget to the headcount budget. That is a ten-times repricing event at the same compute cost. The narrative is not decoration on top of the business model. The narrative is the pricing infrastructure.
Now the technical side, because the number deserves at least one honest stress test before dismissal.
METR's task-horizon research gives us the closest measurable proxy we have: the length of human task a model can complete at 50% reliability has been doubling roughly every seven months. Mid-2025 baseline sat around fifty minutes. Extrapolate eight to nine doublings to 2030 and you land somewhere in the hundreds of hours — weeks of human work per task. That extrapolation does support the claim's order of magnitude. The 75% figure is not fantasy.
But extrapolation hides three fractures. First, the training data wall and diminishing returns on compute. Second, the reliability gap between 50% and 99% — the enterprise adoption threshold — is not a linear climb, it is a cliff with a moat. Third, error compounding: a 99% per-step success rate across a hundred-step task yields 37% end-to-end. Long-horizon autonomy does not degrade gracefully.
Then the finding that should end the conversation: METR's July 2025 randomized controlled trial had experienced open-source developers use AI tooling on real tasks. Actual completion time increased roughly 19%. Their subjective assessment was that they had gotten about 20% faster. Perception and measurement diverged by nearly forty points, in opposite directions.
And every "75% by 2030" figure in circulation comes from exactly that — subjective expectation elicitation. The instrument is known to be miscalibrated in a specific, documented direction. The forecast inherits the calibration error of its method.
There is also an Amdahl's-law-shaped problem the framing elides. Automating 75% of subtasks does not compress 75% of work hours. If the residual 25% — judgment, accountability, cross-party coordination, exception handling — becomes the bottleneck, aggregate time savings collapse toward the serial fraction. Task-level automation rate, hour-level substitution rate, and headcount reduction rate are three different numbers. The discourse treats them as one. That semantic slide is the load-bearing fraud in the entire genre.
Consider also who is being surveyed. Anthropic employees are the most extreme early-adopter sample on earth. Their job is building and consuming frontier AI, with internal agent tooling available to every desk. Extrapolating that cohort to the general workforce is textbook selection bias — not a subtle one, a first-page-one one.
Four probes are missing and the omission benefits the source:
Does 75% mean AI completing work independently, or human-plus-AI completing it? These have opposite labor-market implications. Replacement versus augmentation is not a rounding error, it is the entire policy question.
Is this a capability forecast — what AI will be able to do — or an adoption forecast — what organizations will actually deploy? Historically the lag between the two runs three to five years.
Does Anthropic hold internal telemetry — token consumption, share of commits authored by agents, task completion ratios — that would empirically check the expectation? If internal measurement exists and only the expectation was released, the information value drops by an order of magnitude. Note the release channel: staff expectation, not executive prediction. That is designed deniability. If challenged, it was always an internal survey result.
What is the distribution? A 75% median with wide variance means something completely different from a 75% mean with tight variance. A real internal survey would almost certainly show disagreement. The absence of any dissenting voice is an editorial choice, not a finding.
On labor impact, the honest evidence contradicts the headline in an interesting way. Stanford Digital Economy Lab's 2025 work found employment for 22-to-25-year-olds in AI-exposed occupations declining roughly 13% relative to trend, while senior workers were essentially unaffected. That is not a 75% replacement story. It is a junior-role collapse paired with senior-role amplification — asymmetric, already underway, and needing no wait until 2030.
Which means the actionable forecast is the opposite of the loud one. The displacement is happening at the entry level now. The 75% figure is a distraction that delays preparation by four years.
Sector variance dwarfs any cross-industry mean. Software development is augmentation-dominant with replacement concentrated in templated CRUD and test generation. Content production is replacement-heavy. Customer service is replacement-heavy until compliance and emotional labor intervene. Legal is augmentation-dominant — contract triage automates, liability attribution blocks full autonomy. Clinical work stays under 10% replacement on regulatory barriers alone. A single 75% number averaged across these is not a forecast, it is noise with a headline attached.
And the indirect layer is uncounted. If knowledge-work output costs drop by half, the casualties are not only knowledge workers. It is the service economy orbiting them — consulting, training, recruiting, editing, typesetting, agency work. Those workers do not self-identify as AI-exposed and hold no retraining resources.
Now the counterintuitive part, because the bulls are not simply wrong.
They are right that the capability curve is steep and that their timeframes have beaten consensus for three consecutive years. They are right that enterprise deployment is accelerating and that agent-class workflows are producing real, measurable productivity in narrow domains — code review, test scaffolding, contract first-pass, report assembly. They are right that METR's horizon curve, if it holds, puts weeks-long autonomous tasks within the decade, and that this genuinely does cover most of a knowledge worker's combinatorial task set.
They are right, above all, that the direction is not in dispute.
Where they are wrong is in treating a directional trend as a delivery schedule, and in pricing enterprise contracts against the delivery schedule rather than the trend. The gap between those two readings is where 2027 and 2028 become uncomfortable for everyone holding narrative multiples.
The infrastructure constraint is worth one paragraph because nobody publishes it. If 75% of knowledge work migrates to agent-class inference, the compute requirement is a multiple of current global inference supply — not a percentage increase. Agent sessions consume an order of magnitude more tokens than chat sessions: long context, multi-turn tool calls, retry loops. The load profile shifts from bursty low-concurrency to sustained high-concurrency with persistent state. That demands memory bandwidth and KV-cache economics, not the training-grade interconnects receiving most capital expenditure. Expect the forecast to be vetoed at the power layer before it is validated at the capability layer.
I have spent years pulling apart financial models for a living. The pattern I trust is not the forecast, it is the distance between a forecast and its measurement instrument. When that distance is forty points wide and points in the wrong direction, the forecast is not a prediction. It is a sales artifact wearing a prediction's clothes.
The correct disposition is neither belief nor dismissal. It is classification: treat "75% by 2030" as a sample of vendor narrative strategy — informative about Anthropic's pricing ambitions and media amplification mechanics — and refuse to treat it as a labor-market input. Those are two different objects that happen to share a string of characters.
Three probabilities, stated cleanly so they can be checked later. The probability that Anthropic employees will still report a 75% expectation in 2030 is high, because narrative lock-in and cohort conformity are self-sustaining. The probability that measured AI task-hour share reaches 75% by 2030 is somewhere near 20 to 30%. The probability that headcount shrinks 75% as a consequence is under 5%, because the decay between task automation and establishment reduction is profoundly nonlinear.
What resolves this is not more forecasts. It is a single unglamorous metric: the realized hour-compression rate in enterprise deployments, measured against task-level baselines, tracked across renewals. That number exists inside every enterprise contract signed in the last eighteen months. It will leak within two years. When it does, it will settle an argument that a thousand headlines could not.
Until then, the honest position is that we are pricing a projection whose instrument is documented as miscalibrated, whose sample is the most biased available, whose semantics slide between three different measurements, and whose release channel was chosen for deniability.
That is not a forecast. That is a well-constructed narrative with a number attached.
The compiler is fine. The intent is what needs auditing.