The question everyone asks backwards

"Will AI take my job?" is the wrong question. AI doesn't take jobs — it takes interactions. The right question is: when AI shows up in your work, does it act as an exoskeleton (amplifying what you produce) or as a substitute (doing the task while you supervise it out of existence)?

We built the AI Leverage Index to answer that with observed data instead of vibes. Using task-level AI usage data covering 257 real occupations — the largest public dataset of its kind — we measured the share of each occupation's AI use that is dependent-offload (the human delegates the task and checks the output) versus amplifying (the human stays in the loop doing the judgment work). We then fused that with the structural anatomy of each occupation — how much of the job is information processing, and how much is irreducibly human accountability.

The finding in one line: the jobs where AI is most capable are only dangerous when nobody in the job is accountable for the output.

The most amplified jobs in America are medical

The top of the index is not software. It's the exam room:

OccupationAI LeverageWhy it wins
Neurologists*Top 1%AI processes the imaging and literature; the physician owns the diagnosis
Emergency Medicine Physicians*Top 2%Extreme information throughput, extreme accountability
Nurse Anesthetists*Top 3%Real-time physiological judgment AI can inform but not own
EpidemiologistsTop 3%Highest "AI-as-tutor" pattern in the data — 3 of 4 uses are learning
Obstetricians*Top 5%Physical presence + catastrophic cost of error

* estimated rank — no observed AI-usage data for this occupation; score imputed from structural characteristics (see methodology)

These occupations share an anatomy: heavy information-processing load (which AI is genuinely good at) wrapped around a human accountability core — decisions where being wrong has consequences, so a person must stay on the hook. AI raises their ceiling rather than eroding their floor.

The delegation list is where the story gets uncomfortable

At the other end of the measured data are occupations where over half of observed AI interactions are dependent-offload — the human handing the task over entirely:

OccupationDelegated shareWhat that means
Actors~66%AI writes the lines; the human performs them
Interpreters & Translators~60%Machine translation did the work; humans check register and nuance
Poets, Lyricists & Creative Writers~57%Drafting itself is being offloaded
Special Education Teachers (Preschool)~57%IEPs and documentation — the paperwork, not the classroom
Medical Secretaries~52%Scheduling, records, transcription — near-total offload

Read that list again. It is not "low-skill" work — it contains creative professionals and educated specialists. What unites them is that the accountable artifact is the task itself: a translation, a draft, a schedule. When the artifact can be produced end-to-end, the role compresses to quality control on a shrinking artifact stream.

BLS data is already confirming the split

This isn't a forecast. Between the 2024 and 2025 base-year estimates in BLS's National Employment Matrix, the pattern is already visible in employment counts:

  • Customer Service Representatives: −148,000 jobs (−5.3%) — the single largest occupation-level employment decline in the matrix, and one of the most delegatable information-handling roles
  • Bookkeeping & Auditing Clerks: −81,000 (−5.0%) · Receptionists & Information Clerks: −60,000 (−5.9%) · Office Clerks: −46,000 (−1.7%)
  • Meanwhile the biggest gains: Home Health & Personal Care Aides (+329,000), Registered Nurses (+74,000), Mental Health Counselors (+50,000) — high-accountability, high-contact work

The labor market is quietly repricing the same axis the index measures: delegatable information work is shrinking; accountable human work is growing.

BLS's own forecasters agree. Between their 2024–34 and 2025–35 projections, the Bureau cut its ten-year growth outlook for analyst-heavy work across the board — actuaries from 21.8% to 9.2%, operations research analysts from 21.5% to 11.9%, personal financial advisors from 9.6% to 1.4%. Even the official numbers are re-rating the delegatable tier downward.

The third pattern: AI as tutor

Buried in the data is a third behavior — occupations where the dominant AI use is neither offload nor amplification but learning. Pharmacists (80% of AI interactions are learning-type), exercise trainers, tax preparers, genetic counselors. In these jobs, AI functions as the world's most patient continuing-education department. It is the least reported and arguably most durable pattern: in high-stakes fields, AI that makes the human smarter is worth more than AI that does the task.

What this means for you

Your exposure isn't determined by your industry or your degree — it's determined by the anatomy of your specific role: how much of it is information processing, and who is accountable when the output is wrong. Two people with the same job title can sit on opposite sides of this line.

That's why generic "AI-proof career" lists fail. The question is whether your mix of tasks leans toward judgment or toward artifacts — and that's measurable.

Find out where your work sits on the index

Our free AI Story check scores your field from the same data — amplified, delegated, or barely touched — in about 30 seconds.

Get Your AI Story — Free

Or map your whole work pattern with August's Compass →

Methodology: AI Leverage scores combine observed task-level AI usage patterns (interaction-type shares: directive, feedback-loop, learning) across 257 occupations with structural occupation data from the U.S. Department of Labor's O*NET database — information-processing intensity vs. accountability anchors. Occupations without observed usage are scored by model-based imputation and flagged as such. Employment figures: BLS National Employment Matrix base-year estimates, 2024 vs 2025; growth-outlook revisions from BLS Employment Projections 2024–34 vs 2025–35 releases. This article publishes index scores and rankings only; construct composition is proprietary.