SANTISMM Labs · AI & Work

The professions most exposed to AI

What AI could do versus what it is actually doing — across the 22 occupational groups, with the real-world gap as of June 2026.

Data as of June 2026Anthropic Economic Index · Eloundou et al. · ILO · OECD · IMF

To talk rigorously about “the jobs most affected by AI” you have to separate three things: how many of an occupation's tasks a model could technically do (theoretical capability), how much AI is actually used in that work today (observed coverage), and the net effect on employment and wages. This page reproduces and extends Anthropic's radar of the 22 SOC occupational groups: a large, systematic gap between potential and practice — and what the broader evidence says it means.

20%40%60%80%100%ManagementBusiness& FinanceComputer& MathArchitecture& EngineeringLife & SocialSciencesSocialServicesLegalEducation& LibraryArts & MediaHealthcarePractitionersHealthcareSupportProtectiveServicesFood &ServingGroundsMaintenancePersonal CareSalesOffice& AdminAgricultureConstructionInstallation& RepairProductionTransportation

Hover or tap a category to see its figures.

Anthropic published exact observed values only for the 7 most-exposed groups (plus Architecture & Engineering, derived). The remaining groups are transparent estimates anchored on the qualitative bands in the literature, shown as hollow markers and flagged “estimated”.

Ranking by occupational group

The 22 SOC major groups. Sort by what AI could do (theoretical), what it actually does today (observed), or the gap between them.

Sort by
#GroupTheoreticalObservedGapMechanism
1Computer & Math
94.3%
35.8%
58.5%Complement & speed-up
2Office & Admin
90%
34.3%
55.7%Task substitution
3Business & Finance
94.3%
28.4%
65.9%Complement & speed-up
4Sales
62%
26.9%
35.1%Task substitution
5Legal
89%
20.4%
68.6%Task substitution
6Arts & Media
83.7%
19.2%
64.5%Task substitution
7Education & Library
61.7%
18.2%
43.5%Latent potential
8Management
91.3%
9%*
82.3%Latent potential
9Life & Social Sciences
77%
6%*
71%Latent potential
10Social Services
50.5%
5%*
45.5%Latent potential
11Healthcare Practitioners
59.9%
5%*
54.9%Latent potential
12Architecture & Engineering
84.8%
4.2%
80.6%Latent potential
13Production
19%
2%*
17%Resilient to LLMs
14Healthcare Support
28.5%
1.5%*
27%Resilient to LLMs
15Protective Services
31.6%
1.5%*
30.1%Resilient to LLMs
16Installation & Repair
18.4%
1.5%*
16.9%Resilient to LLMs
17Transportation
12.1%
1.5%*
10.6%Resilient to LLMs
18Food & Serving
16.9%
1%*
15.9%Resilient to LLMs
19Personal Care
18.2%
1%*
17.2%Resilient to LLMs
20Grounds Maintenance
3.9%
0.5%*
3.4%Resilient to LLMs
21Agriculture
15.7%
0.5%*
15.2%Resilient to LLMs
22Construction
16.9%
0.5%*
16.4%Resilient to LLMs

* Estimated observed value (not officially published). The “Gap” is theoretical minus observed, in percentage points: the headroom between what AI could cover and what it currently does.

The single most-exposed occupations

By observed coverage — actual Claude usage. Computer programmers top the list at 74.5%.

  1. 1Computer programmers
    74.5%
  2. 2Customer service representatives
    70.1%
  3. 3Data entry keyers
    67.1%
  4. 4Medical records specialists
    66.7%
  5. 5Market research & marketing analysts
    64.8%
  6. 6Wholesale & manufacturing sales reps
    62.8%
  7. 7Financial & investment analysts
    57.2%
  8. 8Software QA analysts & testers
    51.9%
  9. 9Information security analysts
    48.6%
  10. 10Computer user support specialists
    46.8%

At the other end: around 30% of workers have effectively zero observed coverage — cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers and dressing-room attendants among them.

Augmentation vs. automation

How AI is used in real conversations, not just whether it could be used.

57%
43%
AugmentationAutomation

Anthropic Economic Index (2025). The automation (“directive”) share is rising over time — from 27% (Dec 2024) to 39% (Aug 2025) on Claude.ai, and ~77% on the enterprise API.

Exposure is potential, not destiny

The gap between the blue and red curves — often 50–70 points — means displacement is not imminent: today's use is majority augmentation, and no clear rise in unemployment for exposed workers has appeared yet. The number to watch is whether the red curve climbs toward the blue.

What the evidence actually says

Capability is not adoption.

Knowledge work scores 80–94% on theoretical capability but observed coverage is a fraction of that (Computer & Math: 94% vs 36%). Data governance, legacy systems, liability and human-in-the-loop review slow the move from “can” to “does”.

It's white-collar work this time.

Unlike past automation waves, exposure now concentrates in higher-paid, more-educated office work. Anthropic's most-exposed quartile earns ~47% more and is 16 points more likely to be female — driven by occupational segregation in clerical and administrative roles.

Augmentation still leads — for now.

Real use is ~57% augmentation vs ~43% automation, though the automation share is climbing. The same task can be accelerated by a human or delegated to the model; today the balance still tilts to acceleration.

Watch the youth-hiring canary.

No clear aggregate unemployment effect has appeared, but the monthly job-finding rate for 22–25-year-olds in exposed occupations fell ~14% after ChatGPT — the strongest early-warning signal, echoed by independent research.

Substitution vs reconfiguration.

Clerical, data-entry and first-line support face direct task substitution. Software, analysis, finance and law are mostly being reconfigured — copilot and task delegation, not wholesale replacement — with entry-level rungs being squeezed.

Convergence is a hypothesis.

Anthropic expects observed use to rise toward theoretical capability, but that is an assumption about adoption, not an empirical finding. For many professions the gap may persist far longer than for software.

The macro picture

Headline estimates from the major authorities. These are modeled scenarios with wide uncertainty — not realized outcomes.

80% / 19%
Tasks exposed (US workforce)

~80% of US workers could have ≥10% of their tasks affected by LLMs; ~19% could see ≥50% affected.

OpenAI · UPenn (2023)
≈300M
Full-time jobs exposed globally

Generative AI could expose the equivalent of 300M FTE jobs to automation; up to ~7% potential lift to global GDP.

Goldman Sachs (2023)
40% / 60%
Jobs exposed (global / advanced)

~40% of jobs globally are exposed to AI — about 60% in advanced economies, 26% in low-income.

IMF (2024)
+78M net
Net new jobs by 2030

170M jobs created and 92M displaced by 2030 — a 22% structural churn of the labour market.

WEF (2025)
+$13T
Annual GDP impact by 2030

Up to ~29.5% of US work hours could be automatable by 2030; ~12M US occupational transitions.

McKinsey (2023–24)
1 in 4
Workers with some exposure

1 in 4 workers globally are in an occupation with some GenAI exposure; 3.3% in the highest category, skewed toward women.

ILO (2025)

Sources and methods differ (task rubrics, abilities indices, usage logs), so the figures are not directly comparable and should be read as scenarios.

Methodology & sources

The radar reproduces Figure 2 of Anthropic's “Labor Market Impacts of AI: A New Measure and Early Evidence” (Massenkoff & McCrory, March 2026). The blue curve is the theoretical-capability β from Eloundou et al., “GPTs are GPTs” (2023/2024); the red curve is Anthropic's observed-coverage measure, built from millions of real Claude.ai and API conversations mapped to the ~800 occupations and ~17,000 tasks of the US O*NET database.

Theoretical β is confirmed for all 22 groups. Anthropic published exact observed values only for the seven most-exposed groups; Architecture & Engineering is derived from its stated ~5% observed/theoretical ratio. For the remaining groups we use transparent estimates anchored on the qualitative bands reported across the OECD, ILO, IMF, WEF and McKinsey literature — clearly flagged “estimated” and never presented as official figures.

Because the observed measure is built on Claude data alone, it likely understates real exposure (it misses ChatGPT, Gemini, Copilot and embedded tools), while the theoretical β reflects early-2023 capabilities and probably understates today's frontier. Different methodologies disagree by design: exposure is potential, not displacement.

Primary sources

Compiled by SANTISMM from public research. Figures are rounded; estimated values are flagged. Exposure measures potential task overlap, not realized job loss — every authoritative source stresses transformation over wholesale replacement.