97% vs 47% — Your Job Title Says Nothing About Your AI Exposure
Anthropic's index found 97% of tasks are automation-dominant through the API and 47% in chat. Same model, overlapping tasks. Exposure follows the deployment, not the job title — so stop forecasting the fog and run the migration test instead.
- #career
- #job-market
- #planning
- #international
Toby E. Stuart, a professor at UC Berkeley's Haas School, gave this feeling a name in HBR last April: the fog. His premise, in the part of the piece that isn't paywalled, is that AI has put new limits on how far ahead anyone can see before committing to something long. I could only read the opening, so I won't pretend to summarise his prescription — but the name is right, and the instinct it triggers is wrong. Fog makes you hunt for a better forecast. There isn't one. There is a present-tense measurement almost nobody runs.
The number that kills job-title thinking
Anthropic's Economic Index split its data by where Claude runs. In the September 2025 report, 47% of tasks on Claude.ai — a person in a chat window — showed automation-dominant patterns. Through the enterprise API, that figure was 97%. Same model, overlapping tasks. Business use looked like automation 77% of the time, against roughly 50% for consumer chat.
Read it twice. Whether work is "automated" is not a property of the work, and it is emphatically not a property of your job title. It is a property of the deployment. A task is exposed the moment someone wires it into something that runs with no one watching — and that has either happened or it hasn't. It's observable. It doesn't need a forecast.
The migration test — one hour, every quarter
Forget your job description. List the last ten things you actually shipped, and mark each one:
- Chat-shaped — the input is ambiguous, it needs your judgement mid-flight, and failure is expensive and hard for a machine to notice.
- Pipeline-shaped — structured input, machine-checkable output, and someone could put it on a
crontomorrow.
Every pipeline-shaped item is already sitting on the automation side of that 97%, whether or not your employer has got round to it. The fraction is your exposure number. It moves each quarter, which is the whole point: it's a measurement you repeat, not a prediction you get to be wrong about exactly once.
Durable, or just temporarily scarce?
The other half of planning in fog is knowing what to bet on. The March 2026 "Learning curves" report gives that decision an actual marker. Users with six or more months on Claude had about 10% higher success rates than newcomers — and still 3 to 4 percentage points higher inside identical task categories, so it isn't just veterans picking easier work. The work they bring gets harder too: the education level a task implies rises by almost a year for every additional year of use. And against the obvious prediction, long-tenure users drifted toward validation and learning, not toward full automation.
For Brazilians, the fog has a second layer
If you're in São Paulo aiming at London or New York, you also have a sampling problem. In that same September 2025 data, Brazil was 3.7% of global Claude usage, and per-capita adoption tracks income tightly — 1% higher GDP per capita goes with roughly 0.7% higher usage per person. Israel ran at 7.0x its expected share, Singapore at 4.57x. Brazilian usage skewed toward translation and legal work; India's skewed hard to software.
Your local peer group is not a sample of the market interviewing you. Calibrate off the devs around you and you're reading a distribution that lags the one you're applying into. Pull your signal from the target market — job posts, changelogs, open repos there — not from the room you're in.
Do this before the quarter ends
Run the migration test on your last ten work items. Take the single most pipeline-shaped one and move one level up on it: own the spec, the evaluation, or the failure handling instead of the execution. One task, one level, ninety days — then re-run the count.
My rule: if a career bet can't be checked inside one quarter, it isn't a plan, it's a mood. Optionality isn't a vibe. It's a short list of bets, each with a date you look at it again.
Sources
Keep reading
More guides like this one.
Three Five-Year Plans, or One Plan With Three Salaries?
Stanford's official Odyssey worksheet has one instruction almost everyone skips: make each of the three plans wildly different. Most engineers write three price tiers of the same life instead. Here is the incompatibility test and a prompt that attacks your plans.
Laid Off? These 30 Days Fill Your Calendar, Not Your Résumé
A remote US or EU hiring loop almost never closes inside a month, so a 30-day plan that promises an offer is lying to you. Here is the week-by-week version that ends in a full pipeline instead — and where the free AI credentials honestly belong in it.
Three Moves Into AI Work, Three Different Interviews
Google's own course page says 67% of gen AI users *estimated* it saves them 2+ hours a week. Estimated. Every candidate walks in carrying that same guess, and one real measurement beats all of them. Prep split by which of the three moves you are making, plus a practice-interview prompt that attacks your numbers first.