The YourJobs AI Vulnerability Index is a 0-100 composite score per US occupation that estimates how much of the typical task mix is exposed to current-generation AI within five years. Lower scores indicate roles where physical interaction, accountable judgement, or interpersonal trust dominate. Higher scores indicate roles where a substantial share of tasks is already within AI capability.

It is not a forecast that any specific role will disappear. It is a measurement of capability, not of labour-market outcomes. Whether a role actually shrinks depends on regulation, capital allocation, demand growth, productivity gains being passed to workers, and many other factors the Index does not model. Treat the score as a capability map.

What the bands mean

The 0-100 range is partitioned into four interpretive tiers. The bands are not equal-width; they reflect plateaus where qualitative interpretation changes.

  • Low (0-25): Tasks are largely physical, interpersonal, or judgement-heavy in ways current AI cannot replicate. Examples: registered nurses (18), occupational therapists (14), home health aides (9). The work blends clinical judgement, hands-on procedure, and high-stakes interpersonal trust.
  • Moderate (26-50): Some tasks are AI-augmentable; near-term role evolution rather than replacement. Examples: heavy and tractor-trailer truck drivers (47), assemblers and fabricators (43), machinists (36). Augmentation is real, but autonomous execution still requires human oversight.
  • High (51-75): A significant share of routine work is AI-augmentable; structural pressure on headcount is plausible within five years. Examples: customer service representatives (78 → just over the cutoff into Very High), cashiers (71), administrative assistants (68).
  • Very High (76-100): Most tasks are within current AI capability; aggressive role redesign is expected within three years. The empirical evidence here is thinnest because so few occupations are squarely in this band — the methodology deliberately reserves this tier for high-confidence cases.

How the score is calculated

The composite is a weighted average of three published research signals, each scaled to a common 0-100 range and rounded to integer:

score = round(0.40 × frey_osborne_2013 + 0.40 × openai_penn_2023 + 0.20 × anthropic_economic_index_2024)

Forty per cent goes to Frey & Osborne's 2013 automation-probability table — the most-cited mapping in the field, peer-reviewed in Technological Forecasting and Social Change. Forty per cent goes to OpenAI/Penn's 2023 task-level exposure scores from "GPTs are GPTs" on arXiv. Twenty per cent goes to the Anthropic Economic Index, which measures observed-in-practice Claude usage across occupations.

The weights balance theoretical exposure (Frey & Osborne predates LLMs but covers all 702 SOC occupations rigorously) against LLM-era task analysis (OpenAI/Penn captures what Frey & Osborne missed) against measured workplace adoption (Anthropic's data reflects which roles actually get LLM usage today). For the full disclosure including limitations and citation format, see the methodology page.

How should I use this score?

The single most useful framing is "how much of the typical task mix is at stake?", not "will my specific job disappear?". Two workers in the same SOC occupation can face very different exposure depending on the specific employer, industry, and tools they use day-to-day. The Index captures the typical mix at the occupation level — not your individual situation.

Treat the score as a starting point for three concrete questions:

  1. Which of my tasks would AI struggle with? A Low-band occupation gives the answer at the role level (most of them). A High-band occupation forces you to identify the specific tasks where you add value AI can't replicate.
  2. What does AI augmentation look like in my work? Even Low-band occupations have AI-touched components (documentation, scheduling, knowledge retrieval). Identifying these is how you stay productive as the tools improve.
  3. If I had to pivot, which adjacent occupations would draw on my existing skills? The BLS task-similarity matrix lets you compare task overlap between occupations. A retail salesperson (39) and a customer service rep (78) share many skills but have very different exposure profiles.

Why these particular weights?

The 40/40/20 split is a judgement call we've documented openly. The argument for it: Frey & Osborne's probability column has the broadest coverage and most thorough peer review, so it anchors the scale. OpenAI/Penn corrects for the fact that 2013 didn't have LLMs — adding it at equal weight prevents the composite from underweighting cognitive automation. Anthropic gets a smaller weight because observed Claude usage reflects current adoption, which is partial and skewed toward early adopters; it's a calibration signal, not the centre of gravity.

If you disagree with the weights, the dataset is published under Creative Commons Attribution 4.0: download the components and reweight to your own judgement. We'd genuinely like to hear about alternative weight schemes that fit the data better.

What the Index doesn't tell you

Several things sit outside the model:

  • Wage trajectories. Two roles can have identical AI exposure scores and very different wage outlooks depending on labour supply, regional concentration, and unionisation. The Index doesn't model wages — for that, see the median-wage and projected-growth columns from BLS on each occupation page.
  • Geographic concentration. A Very High exposure score might matter more in a metro where the role is concentrated (e.g. accountants in a financial-services hub) than in one where it's diffuse. We're building a per-metro overlay later this year.
  • Quality of remaining work. AI exposure measures volume of tasks at risk, not whether the residual tasks are intrinsically more or less satisfying. A role could have low exposure and unhappy practitioners; the Index is silent on this.
  • Speed of adoption. The score is "exposure to current-generation AI capability". Whether your specific employer adopts the tools to replace those tasks depends on capital, regulation, and management decisions the data can't see.

Frequently asked

Is a score of 70 the same as "70% chance of being replaced"?

No. The 0-100 range is a relative ranking, not a probability. A 70 means the occupation sits in the High band: a substantial share of routine work is AI-augmentable. It does not mean any specific worker has a 70% chance of losing their job.

Why isn't my exact job title listed?

The Index uses the US Standard Occupational Classification (SOC) 2018, which groups specific job titles into ~800 detailed occupations. We currently cover the most common 35; coverage expands as we map more titles. Use the closest match — your day-to-day tasks usually overlap heavily with the occupation we've scored.

How often do you update?

We refresh the underlying scores annually as new evidence from the source studies is published. The next planned update incorporates the OpenAI/Penn 2026 update and the latest Anthropic Economic Index release. Live job counts on each occupation page refresh hourly from our partner feeds.

What to do with this

If you're a job seeker: use the Index to inform career-pivot decisions. A High-band occupation that overlaps in skill with a Low-band one (e.g. customer service rep → patient care coordinator) is a high-leverage pivot. The methodology page lists the source studies for further reading.

If you're a researcher or journalist: the dataset is freely citable under CC-BY-4.0. We publish the full methodology, components, and limitations on the methodology page. Email the editorial team if you'd like the underlying CSV.

If you're a career counsellor: feel free to use the Index in advising sessions. We'd be interested to hear from you about which framings work and which don't — email [email protected].

The full ranking is sortable and linkable: browse the AI Vulnerability Index.