The market doesn't wait for methodology. It waits for a number. This week the number was 75%. Anthropic staff, per a third-hand brief that surfaced on Crypto Briefing, expect AI to handle 75% of their work by 2030. Three information points. No sample size. No questionnaire wording. No publication date. And yet "autonomous agent" repriced across a dozen tokens before a single reader asked who, exactly, was surveyed. I have spent eighteen months building and backtesting agent-trading bots — I ran a four-developer team through a 35% alpha strategy against classical technical analysis — so I know how this figure gets consumed downstream. Nobody reads it as a survey result. They read it as a total addressable market.
Here is the chain of custody. Anthropic internal expectation, to technology press, to a Crypto Briefing rewrite. A crypto-asset news desk carrying a story with zero crypto elements in the body. That one fact sets the editorial depth: near zero. This is traffic-driven syndication, not specialist reporting. Every hop strips a qualifier. By the third hop, "a subset of employees in an internal survey" becomes "Anthropic says AI will do 75% of the work."
Why should a chain care? Because this number is the hidden denominator under every AI-agent token multiple. Global software spend is roughly a one-trillion-dollar pool. Global knowledge-work payroll is thirty to forty trillion. Same technology, two TAMs, one order of magnitude of valuation gap. Any narrative that migrates the denominator from the first pool to the second is worth more than any product launch, and it costs nothing to distribute. That is the machine I trade around, and it is running at full throttle.
Now the technical work. Strip the figure to its mechanics and three separate claims are fused into one. Task automation rate is not hours displaced. Hours displaced is not headcount reduction. A survey measuring the first gets quoted as the third, and the gap between them is where fortunes are made and lost. Amdahl's law is unforgiving here: automate 75% of the subtasks and the remaining 25% — judgment, accountability, cross-party coordination, exception handling — becomes the bottleneck. Overall time compression lands far below 75%.
Start with what is measurable. The most relevant proxy is task-time horizon. METR's 2025 work puts the doubling period near seven months, against a baseline of roughly fifty-minute human tasks. Extrapolate eight or nine doublings to 2030 and you reach tasks spanning hundreds of hours — weeks of work. On that curve, 75% is not fantasy. The direction is defensible; the precision is not.
Extrapolation carries three known fractures. The data wall and diminishing compute returns. The reliability cliff — moving from 50% to 99% reliability, the enterprise floor, is not linear effort. And error compounding: a 99% per-step success rate across a hundred-step task leaves 37% end-to-end. The last mile is not a mile. It is a cliff.
Here is the evidence that should stop you cold. METR's July 2025 randomized controlled trial found experienced open-source developers using AI tools finished tasks about 19% slower — while reporting they felt 20% faster. Perceived efficiency and actual efficiency diverged in opposite directions. And the 75% figure comes from the identical instrument: a subjective expectation survey. You are being handed a sentiment reading and asked to price it as a measurement.
The labor market already sketched the real shape, and it is not "75% across the board." Stanford's Digital Economy Lab in 2025 found employment for 22-to-25-year-olds in AI-exposed occupations down roughly 13% relative, while senior workers were untouched. Junior collapse, senior augmentation. That asymmetry is live now, years before 2030. "75%" as a cross-industry average is meaningless, because inter-sector variance dwarfs the mean. Software is augmentation with templated work replaced. Content is high substitution. Legal is augmentation with liability blocking full automation. Medical sits under 10% behind regulatory walls. One number cannot hold all of that.
Add the sampling problem. Anthropic staff are the most extreme early-adopter cohort on earth — their job is to build and use AI, fully provisioned with frontier models and agent tooling. Extrapolating them to society is textbook selection bias. And 75% most likely measures task-level automation, not job-level replacement, a semantic slide the headline completes for free. Note also what was withheld: if internal token-consumption or AI-generated-code share data exists and only the "expectation" was released, the information value collapses.
The brief's one honest word was "disparity" — and it never quantified it. That gap is the real story. Insiders and the public are diverging by tens of percentage points on the same curve. Executives who sign procurement contracts believe the replacement story; the workers who absorb it do not. That contradiction is the primary friction of enterprise AI deployment through 2028, and it is a chain-adjacent opportunity, not a forecast: whoever sells the measurement layer between belief and reality captures the spread.
Then translate to the chain, where I actually work. Agent workloads do not burn tokens like chatbots. Multi-turn tool calls, long context, failure retries — a single agent session can consume tens of times the tokens of a chat exchange. The tell is in the load profile: chat tokens are bursty and cheap to batch; agent tokens are long-lived, stateful, and expensive to keep warm. Datacenter capex is still skewed toward training-class clusters, while agent inference rewards memory capacity and cache hit rates. If even a fraction of that 75% migrates to autonomous execution, the inference bill is not "more GPUs." It is a different infrastructure shape: memory bandwidth, KV-cache economics, sandboxed orchestration, low-latency reliability tiers. The AI-agent tokens I track are pricing that future load today. Speed is currency, but precision is the vault — and right now the vault is being filled with a sentiment number.
And the physical ceiling is rarely computed. Delivering 75% of knowledge work as agent inference implies an electricity and chip bill that could be a multiple of today's global inference supply. If power and accelerators cannot scale to meet it, the expectation dies at the physical layer before it ever dies at the capability layer. That is the physical constraint no one in the headline bothered to price.
Here is the angle nobody reported. The leak architecture is the signal, not the number. Anthropic let "employee expectation" surface, not "executive forecast." If challenged, it reverts to individual opinion inside an internal survey — engineered plausible deniability. That is not an accident. It is how a frontier lab releases a capability claim while keeping an exit. The market read the headline and skipped the grammar.
The grammar is where the trade lives. If the 75% narrative is falsified, it breaks at the enterprise renewal window — 2027 into 2028 — when net revenue retention on agent products either validates outcome-based pricing or collapses back to tool-based pricing. That is a one-order-of-magnitude repricing of the denominator. Crypto AI tokens, all reflexivity and leverage with no earnings to anchor them, will move first and hardest in both directions. When the data lands, the pivot is not a retreat, it is a recalibration — from narrative to measurement. The crowd is positioning for the forecast; the desk is positioning for the falsification.
Watch the instruments, not the predictions. Track the Anthropic Economic Index split between directive and feedback conversations — automation versus augmentation. Track AI-generated code merge and rework rates, the earliest honest read on the last mile. When the survey's source, sample, and wording finally surface — or never do — you will know whether this is data or doctrine. The market doesn't care about your sentiment; it cares about your liquidity. Position for measurement, not for the forecast.