SPB Git

spb/airiskindex Public

The most methodologically rigorous, fully transparent AI job-exposure index.

🔗 https://www.airiskindex.io
webappaiindexlaborresearch
TypeScript 88% Python 6.1% SQL 2.7% CSS 1.2% JavaScript 0.9% Shell 0.8%
27.2 KB
View source Raw Blame History

# 04 — Competitive Landscape & Real-World Evidence (through August 2026)

Research memo for the AI Risk Index (airiskindex.io). Compiled 2026-08-05 from live web research. Purpose: (a) map existing public AI-job-risk tools, (b) assemble the 2024–2026 empirical record on AI's labor-market effects, and (c) propose how this evidence parameterizes the barriers and adoption_velocity scoring dimensions (weights 0.20 / 0.10 in v1, packages/scoring/src/weights.ts).

# Part A — Competitive landscape of public AI-job-risk tools

# A.1 willrobotstakemyjob.com (the incumbent)

  • Since 2017. Tagline: "Find out how likely your job is to be automated — based on real data and user votes."
  • Methodology: Automation-risk probabilities produced "using a similar method" to Frey & Osborne (2013) ("The Future of Employment", Gaussian process classifier over 702 occupations, the famous "47% of US employment at risk" paper), re-estimated "with the most up-to-date data available", plus BLS occupation data (employment, wages, growth). Since 2019 it collects user poll votes on perceived risk; in 2021 it blended BLS data + polls + automation probability into a 0–10 "job score". Source: https://willrobotstakemyjob.com/about
  • Critical weakness: the underlying model is pre-LLM computerization/robotics-era work. Frey–Osborne scored whole occupations on physical/routine automatability (manual dexterity, cramped workspace, fine arts, social perceptiveness...), which inverts under generative AI: it rates cognitive/office work as relatively safe and misses exactly the exposure the 2024–2026 evidence shows (translators, writers, customer service, junior developers). Single doom-number framing ("X% probability of automation"), no exposure/substitution/augmentation distinction, no confidence intervals, no versioning, no API.
  • Traffic (why it still matters): ~71.3K visits in May 2026 per Semrush (down ~20% MoM), world rank ~#50,395; ~42% of traffic from Google organic, ~39% direct. Other estimators give 45K–100K+/month. Sources: https://www.semrush.com/website/willrobotstakemyjob.com/overview/ , https://hypestat.com/info/willrobotstakemyjob.com
  • Monetization: ads; no API or paid tier found.

# A.2 Other public lookup tools (US/EN)

Tool Data / method Output Weaknesses Pricing/API
willrobotstakemyjob.com (2017) Frey–Osborne-style re-estimation + BLS + user polls Single automation % + 0–10 "job score" Pre-LLM model, whole-occupation doom number, no sub-scores, no versioning Free, ads; no API
replacedbyrobot.info ("2026 AI Automation Risk Database") Claims BLS + O*NET over "57,000+ occupations" (job titles, not SOC codes) "2 risk scores per job" (AI + robotics) Opaque method, ad-heavy, title-level pseudo-precision Free, AdSense; no API
aijobimpactcalculator.com (2026, by Digital Signet) ILO 2025 GenAI exposure gradient (4 bands) + Brookings 2024 task rubric on O*NET 30.2 tasks + BLS EP 2024-34 + WEF FoJ 2025; static, pre-computed; published methodology & revision history Exposure band + top-5 tasks tagged Displaceable/Changing/Growing + "what's growing" panel Band-level only (4 bands), no composite score, no CI, US-centric Free; no API
replacemeter.com Undisclosed (looks LLM-generated per submitted job title) Letter grades: "AI Resilience" %, "Adaptability" % No methodology page, arbitrary-title scoring, no provenance Free
tripleten.com/tools/what-jobs-will-ai-replace LLM analysis of user-entered title/industry; "most recent AI research data" (uncited) 0–100% automation risk + skills advice + career alternatives Lead-gen for a bootcamp; unreproducible Free (lead-gen)
ailayoffs.live Aggregates layoffs.fyi, Goldman, McKinsey, WEF; "Oxford research + real layoff data" risk checker Live displacement counters + risk score Doom-counter framing, mixes projections with counts Free
ailayofftracker.com / founderreports.com/ai-layoffs-tracker / skillsyncer.com/layoffs-tracker Curated AI-cited layoff announcements (Challenger, TechCrunch sourcing) Event lists, totals Event trackers, not occupation scores Free
techjacksolutions.com/job-displacement-trends Mash-up: Anthropic Economic Index %, Gartner, BLS growth, "WifiTalents" Per-occupation risk ranges (e.g., customer service 67–80%) Mixes usage shares with risk %, low-quality sources alongside good ones Free
Stanford Canaries Dashboard (digitaleconomy.stanford.edu) ADP payroll microdata, 4.6M workers, 730+ occupations, continuously updated Employment trends by age × AI exposure Research dashboard, not per-occupation risk lookup — but the credibility benchmark Free

Academic/institutional exposure indices that power many of these (no consumer UI of their own): Felten–Raj–Seamans AIOE (AI application ↔ 52 O*NET abilities; https://sites.bu.edu/tpri/2021/06/02/occupational-industry-and-geographic-exposure-to-artificial-intelligence-a-novel-dataset-and-its-potential-uses), ILO Global Index of Occupational Exposure to GenAI (ISCO-08, 4 gradient bands; https://webapps.ilo.org/static/english/intserv/working-papers/wp140/index.html), Microsoft "AI applicability score" (Tomlinson et al. 2025, 200K Copilot conversations mapped to O*NET work activities — top: interpreters/translators (98% activity overlap), historians, passenger attendants, sales reps, writers, customer service reps; https://www.microsoft.com/en-us/research/publication/working-with-ai-measuring-the-occupational-implications-of-generative-ai), OpenAI "GPTs are GPTs" (Eloundou et al., Science 2024), Pew 2023 O*NET work-activity classification (https://www.pewresearch.org/social-trends/2023/07/26/2023-ai-and-jobs-methodology-for-onet-analysis).

# A.3 French / EU equivalents

  • jobimpact.aidoption.fr — "Exposition IA du marché de l'emploi français": 532 ROME occupations, treemap (surface = jobs, color = 0–10 exposure), France Travail ROME + DARES data, scored by Claude (LLM-as-rater — directly comparable to our §6 pipeline, but with no audit trail or versioning). Has a /us/ twin.
  • transitions-ia.fr ("IA & Métiers France" observatory) — ROME 4.0 + INSEE EEC 2024 + DARES BMO 2024; task-level exposure score per métier. https://otakuch.github.io/transitions-ia.fr/
  • job-guard.com — French-language "votre métier va-t-il disparaître ?" test; editorial/affiliate quality.
  • Observatoire des Emplois Menacés et Émergents + Coface study (Nov 2025, covered by Les Échos): ~16% of French jobs at risk; white-collar metropolitan jobs most exposed (Paris ~19%, Lyon/Toulouse 18%); legal/accounting, publishing/press, IT programming/consulting, insurance, finance >25% of jobs exposed. https://www.lesechos.fr/monde/europe/ia-le-grand-bouleversement-a-venir-du-marche-du-travail-2221760
  • No credible official FR/EU consumer lookup exists (France Travail offers only e-learning content) → the ESCO/ROME crosswalk in our roadmap targets an empty niche.

# A.4 SEO landscape

  • Queries like "will AI take my job (2026)" are dominated not by tools but by listicle/content marketing: Nucamp, NovoResume, Careerminds, AI Weekly, Shawn Kanungo — all citing the same WEF 92M-displaced / 170M-created / +78M-net figure, plus Goldman's −16K net jobs/month (Apr 2026). Tool sites rank on "AI job risk calculator" / "will robots take my job" variants; willrobotstakemyjob.com still owns its brand query with ~42% organic share of its traffic.
  • Opportunity: nothing ranking today combines (1) task-level methodology, (2) sub-scores with uncertainty, (3) live evidence (Challenger/adoption data), (4) versioned transparency, (5) non-doom adaptation framing. aijobimpactcalculator.com is the closest philosophical competitor (source-cited, anti-doom, "what's growing" panel, "how to argue with this" page) but is static, band-level, and has no composite index, no API, no EU coverage.

# Part B — Real-world evidence, 2024 → August 2026

# B.1 Entry-level employment effects ("Canaries" line of evidence)

# B.2 Layoffs attributed to AI (Challenger, Gray & Christmas)

# B.3 Sector-specific substitution evidence

# B.4 Adoption statistics (the adoption_velocity evidence base)

  • Census BTOS (official, firm-weighted): AI use in core production 3.8% (Sep 2023) → 4.6% (early 2024) → ~10% (Sep–late 2025, doubling in ~18 months). Question broadened Nov 2025 to "any business function": 17–20% of firms Dec 2025–May 2026; 20–23% expect use within 6 months. https://www.census.gov/library/stories/2026/05/ai-use-businesses.html
  • BTOS AI Supplement (Nov 2025–Jan 2026): 18% of firms firm-weighted = 32% employment-weighted; very large firms in Information / Professional Services / Finance: 50–60% (60–70% employment-weighted); 57% of adopters use AI in ≤3 business functions (top: sales & marketing 52%, strategy/biz-dev 45%, IT 41%); workers use AI in tasks at 23% of firms (41% employment-weighted). https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-25.html
  • Fed monitoring note (Apr 2026) reconciles the scales: BTOS firms 18%; employment-weighted SBU 78%; 41% of the labor force uses GenAI for work (RPS, Nov 2025; +9.7pp YoY); 50% uses it outside work. https://www.federalreserve.gov/econres/notes/feds-notes/monitoring-ai-adoption-in-the-u-s-economy-20260403.html
  • Ramp AI Index (paid adoption, 70K+ firms' card/bill spend): businesses paying for AI crossed 50.4% in March 2026 (35% a year earlier; 46.8% Jan 2026); jump driven by "late majority" manufacturing/retail; VC-backed startups ~80%. Vendor race: OpenAI 35.2% vs Anthropic 30.6% of businesses (Apr 2026). https://ramp.com/data/ai-index , https://ramp.com/data/april-2026-ai-index
  • McKinsey State of AI (Nov 2025): 88% of orgs use AI in ≥1 function, but ~two-thirds still piloting; 62% experimenting with agents, 23% scaling agents in ≥1 function, ≤10% scaling within a single function; only 39% report any enterprise EBIT impact; ~6% are "high performers" (>5% EBIT from AI). https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
  • Agentic reality check: MIT NANDA: 95% of GenAI pilots show no P&L impact; Gartner: >40% of agentic projects to be cancelled by end-2027; S&P Global/McKinsey: 31% of enterprises run ≥1 agent in production — banking/insurance 47% vs healthcare 18% and government 14%. Gartner: 80% of enterprise apps shipped in Q1 2026 embed an agent (vs 33% in 2024).
  • France anchor: 35% of French firms >10 employees use AI (OPIEC 2025) — EU lags US enterprise adoption.

# B.5 Wage effects & augmentation-vs-substitution

  • Humlum & Vestergaard (NBER w33777; Denmark, 25,000 workers × 7,000 workplaces, 11 exposed occupations, admin-linked): AI chatbots → precisely-estimated null on earnings and hours (CIs rule out >1–2% average effects; occupation-level >6%); avg time savings only 2.8–3%; >80% of saved time reallocated to other work; 8.4% of workers gained NEW tasks created by AI (e.g., checking AI output); minimal pass-through of gains to wages. https://www.nber.org/papers/w33777 , https://www.andershumlum.com/s/chatbots_july25.pdf → Through 2024, within-worker wage effects ≈ 0; displacement shows up at hiring margins first (B.1).
  • PwC Global AI Jobs Barometer 2026 (1B+ job ads, 27 countries): 62% average wage premium for jobs requiring AI skills (118% in consumer markets, 16% in government); AI-skill jobs growing 69% vs 9% market; firms most able to use AI: headcount growth 53% vs 36% and wage growth 24% vs 17% vs least-exposed — "two distinct labour-market paths". https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html
  • Anthropic Economic Index (Claude usage, task-mapped to O*NET): initial report ~57% augmentation / 43% automation; by Nov 2025, Claude.ai = 52% augmentation / 45% automation (automation briefly led mid-2025); enterprise API traffic is automation-dominant (back-office workflows: email, document processing, CRM, scheduling). Directive/full-delegation usage rising. Reports: https://www.anthropic.com/research/anthropic-economic-index-january-2026-report , https://www.anthropic.com/research/economic-index-march-2026-report , https://www.anthropic.com/research/economic-index-june-2026-report → Validates keeping automation-vs-augmentation as the substitution discriminator (Canaries fact: employment declines concentrate in automation-dominant occupations).
  • BCG (Jan 2026, Revelio 1,500 roles): point estimate 10–15% of US jobs vulnerable over 4–5 years; distinguishes substituted vs "divergent" (demand-expansion offsets 0.5–1.0) vs amplified roles.

# B.6 Regulation & barriers (the barriers evidence base)

  • EU AI Act — timeline moved under our feet: GPAI obligations applied Aug 2, 2025. The Annex III high-risk obligations (incl. ALL employment/HR AI: CV screening, targeted job ads, promotion/termination decisions, worker monitoring — Annex III pt. 4) were due Aug 2, 2026 but the "Digital Omnibus" political agreement postpones them to Dec 2, 2027. Requirements when live: risk management, bias testing, logging, human oversight, conformity assessment. Net effect: EU employment-AI adoption friction persists but the binding date slipped ~16 months (a barriers-lowering event for 2026–27 velocity in the EU). https://ogletree.com/insights-resources/blog-posts/eu-nears-approval-of-agreement-to-delay-rules-for-ai-use-in-employment-decisions/ , https://accessfinancial.com/eu-ai-act-recruitment-high-risk-hiring-2026/
  • US states: Illinois HB 3773 effective Jan 1, 2026 (AI discrimination in employment = civil-rights violation; notice required; zip-code-proxy ban). Colorado's AI Act (SB 24-205) delayed to Jun 30, 2026, then replaced by narrower SB 26-189 (May 2026). NYC Local Law 144 (bias audits for hiring tools) ongoing. Pattern: US regulates AI deciding about workers, not AI replacing workers. https://ogletree.com/insights-resources/blog-posts/illinois-steps-up-ai-regulation-in-employment-key-takeaways-for-employers/
  • Professional licensing & liability as adoption brakes (occupation-level, measurable):
    • Legal: AI can't be licensed/disbarred/sworn; legal-specific AI tools show 17–34% error rates; 700+ court cases worldwide involve AI hallucinations with sanctions → yet legal-professional GenAI adoption still jumped 31% → 69% between 2025 and 2026 (54% of firms give no training, 43% no policy). https://www.americanbar.org/groups/law_practice/resources/law-technology-today/2026/whats-really-holding-law-firms-back-from-embracing-ai/
    • Healthcare: litigious + risk-averse, FDA draft guidance (Jan 2025) on AI credibility in drug decisions; physician AI utilization nonetheless 38% (2023) → 72% (2026), concentrated in administrative tasks — augmentation inside a licensing moat.
    • Customer-facing quality floors: Klarna reversal (B.2) = empirical "human-contact requirement" barrier.
  • Barrier taxonomy the evidence supports: (1) statutory/licensing monopoly on the task; (2) liability & error-cost asymmetry (hallucination sanctions, malpractice); (3) regulated-process requirements (EU AI Act Annex III, state notice/audit laws); (4) human-contact/quality preference (Klarna); (5) organizational friction (MIT 95% pilot failure; McKinsey: only 6% high performers; workflow redesign is the differentiator).

# B.7 Sector adoption-velocity differentials (who's fast, who's slow, why)

Sector Signal (2025–26) Why
Information / Tech BTOS large-firm use 50–60%; >90% of tech companies use AI in ≥1 function Digital-native tasks, no licensing, in-house skills
Finance & insurance BTOS top-3 sector; agents in production: banking/insurance 47% (S&P); NVIDIA State of AI: strongest ROI Structured data, measurable use cases; regulation shapes but doesn't block
Professional services (legal, consulting, accounting) Legal GenAI 31%→69% in one year High exposure; liability slows delegation, not use
Telecom / Retail & CPG Agentic adoption 48% / 47% (NVIDIA) Customer-ops scale economics
Manufacturing / logistics Fastest %-growth in AI job postings; Ramp "late majority" surge drove the 50% crossing (Mar 2026) Started low; predictive maintenance, quality control
Healthcare Physician use 38%→72% but agents-in-production only 18%; admin-first Licensing, liability, FDA; augmentation-dominant
Government / public sector Agents 14%; lowest AI wage premium (16%) Procurement, accountability, unionization
Construction / trades ~1.4% adoption (laggard anecdote); BTOS small-firm use <20% Physical, unstructured, small-firm dominated

Cross-cutting velocity facts: firm size is the strongest adoption predictor (BTOS: <20% for ≤4-employee firms); employment-weighted adoption ≈ 2× firm-weighted; paid adoption (Ramp 50.4%) runs far ahead of production-grade deployment (31% ≥1 agent in production) which runs ahead of measured P&L impact (39% any EBIT effect).

# Inputs for barriers & adoption_velocity scoring

# adoption_velocity (weight 0.10) — proposed parameterization

Score each occupation's dominant sector(s) (via BLS OES industry-occupation matrix) on observable adoption, not vendor hype. Candidate sub-indicators, each normalizable to 0–1 with a public, refreshable source:

  1. Sector AI-use rate, employment-weighted — BTOS bi-weekly sector series + AI Supplement (CES-WP-26-25). Anchors: Information ≈ 0.9 · construction ≈ 0.1.
  2. Sector adoption momentum — 12-month delta in BTOS use rate and/or Ramp AI Index sector series (captures manufacturing/retail late-majority acceleration, Mar 2026).
  3. Agentic deployment depth — % of sector firms with agents in production (S&P Global/McKinsey: banking 47% … government 14%). Agents, not chatbots, are the substitution-relevant margin.
  4. Realized displacement intensity — Challenger AI-cited cuts by industry, trailing 12m, scaled by sector employment (chemicals, fintech, tech score high in 2026); optionally corroborated by Canaries Dashboard 22–25 employment trend for the occupation itself.
  5. Occupation-level usage intensity — Anthropic Economic Index share of usage mapped to the occupation's O*NET tasks, split augmentation vs automation (automation share feeds substitution, not just velocity).

Suggested v1 formula: adoption_velocity = 0.35·(1) + 0.20·(2) + 0.20·(3) + 0.15·(4) + 0.10·(5), with the caveat documented in METHODOLOGY.md that (1)–(3) are sector-level priors and (4)–(5) are occupation-level correctors. Update cadence: quarterly (BTOS bi-weekly, Ramp monthly, Challenger monthly, AEI ~quarterly) — each refresh = new INDEX_VERSION patch run.

# barriers (weight 0.20) — proposed parameterization

Score as adoption friction (high barriers ⇒ lower net risk), with five components:

  1. Licensing/authorization requirement (0–1): does task sign-off legally require a licensed human (law, medicine, engineering PE, aviation, finance advice)? Source: O*NET Job Zone + state licensing DBs. Evidence: legal/medical augment-don't-substitute pattern (B.6).
  2. Liability & error-cost asymmetry (0–1): cost of a wrong AI output (malpractice, sanctions, safety). Proxy: occupation's litigation exposure + documented AI-error sanction record (700+ hallucination cases).
  3. Regulatory-process coverage (0–1): is the occupation's automation itself regulated? EU AI Act Annex III (now Dec 2027), Illinois HB 3773, Colorado SB 26-189, NYC LL144 — maintain a dated rule table per jurisdiction; this component is jurisdiction-specific (US vs EU scores diverge; the Omnibus delay is a worked example of a barrier score dropping between index versions).
  4. Human-contact requirement (0–1): reuse O*NET work-context variables ("contact with others", "deal with external customers", physical proximity). Empirical anchor: Klarna reversal; CSAT floors.
  5. Organizational/implementation friction (0–1): sector pilot-failure and scaling rates (MIT 95%, McKinsey 6% high performers, Gartner >40% agent-project cancellations) — a global dampener that decays over index versions as deployment matures.

Suggested v1 formula: barriers = 0.30·licensing + 0.25·liability + 0.20·regulatory + 0.15·human_contact + 0.10·org_friction.

# Positioning implications (for public copy, "adaptation not doom")

  • The 2026 record supports differentiated, hedged claims: entry-level exposure is real and measured (−16% relative, Stanford/ADP; Challenger 101,743 H1-2026 AI-cited cuts) while incumbent wage/hours effects are so far null (Denmark) and AI-skill premia are large (+62%, PwC). That is our three-concept split — exposure ≠ substitution ≠ augmentation — now empirically vindicated; no competitor surfaces it.
  • Publish a "what would change this score" section per occupation (aijobimpactcalculator.com's "how to argue with this" page is the only competitor doing epistemic honesty — match and exceed it).
  • Every headline stat above has a reversal or caveat attached (Klarna, LinkedIn attribution caution, Brookings "no apocalypse yet") — cite these in-product to keep the tone guide credible.