The most useful question about AI and work isn’t “will my job survive?” It’s “which parts of my job survive?” That reframing — from occupation to task — is quietly becoming the consensus method across the people actually measuring this, and three independent efforts now point at the same conclusion from three different directions.
NextGig’s “How AI Changes Your Job” makes the argument visually: every occupation is a bundle of tasks, and AI hits those tasks unevenly. Some get automated, some get amplified, some stay stubbornly human. What makes the tool worth taking seriously is that almost none of it is opinion. The market layer — growth, openings, pay, education — comes straight from the U.S. Bureau of Labor Statistics Employment Projections for 2024–2034. The task layer is lifted verbatim from the Department of Labor’s O*NET database, the same task statements government analysts use to define roughly 800 occupations. The only assessed layer is the AI effect on each task, sorted under a published rubric: automate (software can do it end to end), augment (AI does much of it but a human directs and is accountable), or human (needs physical presence, dexterity, real-time trust, or legal accountability). Across 832 occupations, that’s a lot of judgment calls — but they’re judgment calls applied to a fixed, public inventory of tasks rather than vibes about whole careers.
That task-first approach isn’t unique to NextGig. It’s the same move two of the most-cited studies in this space made first.
The exposure view: how much could be affected
In “GPTs are GPTs,” researchers at OpenAI and the University of Pennsylvania (later published in Science) built their estimate the same way — on top of O*NET tasks — and asked a narrower question: what share of an occupation’s tasks could an LLM make meaningfully faster? Their headline finding is the number that gets quoted everywhere: roughly 80% of U.S. workers are in occupations where at least 10% of tasks are exposed to LLMs, while about 19% are in occupations where more than half the tasks are exposed.
The detail underneath the headline matters more. Looking at LLMs alone, the authors estimated only about 15% of tasks are directly exposed, and just 1.8% of jobs would have over half their tasks affected. The number explodes once you assume the surrounding software — the tools, integrations, and interfaces built on top of the raw model — keeps maturing: under that scenario, the share of jobs with more than half their tasks exposed jumps to roughly 46%. The lesson is that the model is rarely the bottleneck; the scaffolding around it is. Notably, exposure climbed all wage levels rather than concentrating at the bottom, with higher-income work often more exposed — the opposite of the pattern earlier waves of automation followed.
The key caveat the authors themselves flag: exposure is theoretical capability, not a prediction. It says a task could be sped up, not that anyone is doing it, and it attaches no timeline.
The usage view: how much is actually happening
That gap — between what’s possible and what’s real — is exactly what Anthropic’s Economic Index measures, by analyzing millions of anonymized conversations on Claude and classifying each against the same O*NET task list. Three findings stand out.
First, the split between collaboration and replacement leans toward collaboration. In the first report (early 2025), about 57% of task interactions were augmentation versus 43% automation. Across later reports that ratio has wobbled and the automation share has crept up slowly, but augmentation has stayed in the lead — roughly 52–55% in the most recent samples. AI is, for now, more often a thinking partner than a substitute.
Second, adoption is wide but shallow. By the 2026 reports, around 49% of jobs had seen at least a quarter of their tasks touched by Claude — but very few jobs were using it for the majority of their tasks. Usage also clusters hard in one place: software engineering and adjacent computer-and-mathematical work accounted for the largest share of queries (about 37% in the first report). Meanwhile both the lowest-paying and the very-highest-paying occupations showed low AI use, because both lean on physical presence and manual dexterity — Anthropic’s own examples were shampooers and obstetricians. That’s the same “stays human” bucket NextGig draws by hand.
Third, and most revealing: actual usage runs well below theoretical exposure. When Anthropic lined up the OpenAI exposure scores against real Claude usage occupation by occupation, most occupations sat below the line of equality — exposed far more than they were actually being used. In Anthropic’s own capability framework, the tasks rated fully feasible for an LLM made up about 68% of observed usage, while genuinely infeasible tasks were just 3% — so usage concentrates exactly where the model is strong, but the total volume still trails what’s theoretically possible.
Reading your own job
Put the three together and a consistent picture emerges. Exposure (OpenAI) tells you the ceiling: how much of a role could be touched, and it’s a lot, across every income band. Usage (Anthropic) tells you the floor: what’s diffusing today, which is far less, heavily weighted toward knowledge work, and still mostly augmentation rather than replacement. NextGig sits in between, translating both into a per-task verdict you can actually look up for your own occupation.
The practical implication for anyone trying to AI-proof a career is that “job” is the wrong unit of analysis. A role that is 70% automatable tasks and 30% human ones doesn’t vanish — it gets restructured, and the human 30% becomes the part that defines and compensates the work. The tasks that consistently land in the “human” column across all three frameworks are the same ones: physical presence, real-time trust, dexterity, and legal accountability. Those are worth leaning into. The tasks that land in “automate” — drafting, summarizing, routine analysis, code scaffolding — are worth getting very good at directing rather than performing.
A few honest limits. None of this forecasts timing; exposure and usage both describe the present, not a date. The AI-effect labels are assessments, not measurements, even when the tasks underneath them are real. Each source carries its own slant — usage data from a single model’s users skews toward early-adopting, technical professions, and exposure studies depend on a rubric reasonable people can argue with. And the whole frame deliberately brackets robotics: these tools assess what software can do to information tasks, which is why physical work keeps reading as “safe” even where the economics might eventually say otherwise.
But the convergence is the story. Three groups with different incentives, different methods, and different data — government employment statistics, a capability audit, and live usage logs — all decided that the only sane way to reason about AI and work is one task at a time. Your job isn’t a single bet. It’s a portfolio, and the parts are repricing at very different rates.
Sources
- NextGig — “How AI Changes Your Job.” Task-level AI-effect classification built on BLS Employment Projections 2024–2034 (market data) and O*NET task statements. nextgig.rocks/dash/how-ai-changes-jobs
- Eloundou, Manning, Mishkin & Rock — “GPTs are GPTs.” OpenAI/UPenn, 2023; published in Science, 2024. Task-exposure estimates over O*NET occupations. arxiv.org/abs/2303.10130
Anthropic Economic Index. Augmentation-vs-automation splits, adoption breadth, and exposure-vs-usage comparisons from anonymized Claude usage analyzed against O*NET tasks. anthropic.com/news/the-anthropic-economic-index
