Methodology

How we calculate AI job exposure

The AI Task Exposure Score is a transparent, published heuristic — not a black box and not a prediction of job loss. Here is exactly how it works, what sources it rests on, and what it cannot know.

Last reviewed 2026-06-14

01

We score tasks, not people

A job is a bundle of tasks. We estimate exposure at the task level and roll it up — so the result shows which parts of the work are exposed, not a verdict on you.

02

Why job titles are imperfect

The same title means different work at different companies. A title is a starting point; pasting your actual responsibilities (a future feature) will always be more accurate.

03

Why high exposure is not guaranteed replacement

Exposure measures how well AI fits a task — not whether it will be adopted, allowed, trusted, or cheaper than a person. Regulation, cost, culture, and accountability all slow real-world replacement.

04

How we map titles to occupations

We normalise the title (removing seniority words, expanding abbreviations) and match it to the closest O*NET / ESCO occupation and its alternate labels, then attach a confidence level to the match.

Step 5

How we classify tasks

Automatable now

AI can already do most of this task.

AI-assisted now

AI speeds this up but you stay in the loop.

Hard to automate

Needs human judgment; AI only supports.

Human-critical

Depends on accountability and trust AI cannot hold.

Step 6

How we calculate the score

Each occupation is rated 0–100 on eight factors. Five raise exposure; three lower it. The weighted result is normalised to 0–100.

Digital work dependency+15%

Can the task happen mostly through software, documents, spreadsheets, code, or screens?

Language & information intensity+20%

Does it involve reading, writing, summarising, explaining, coding, or communicating information?

Routine & repeatability+15%

Is it pattern-based, rules-based, template-based, or frequently repeated?

Current AI capability fit+20%

Can today's generative AI perform, draft, classify, retrieve, or accelerate it?

Real-world AI usage signal+10%

Do usage studies show AI actually being used for similar activities?

Human judgment & accountability10%

Does it need ethics, high-stakes decisions, taste, or ownership of outcomes?

Physical-world dependency10%

Does it need physical presence, dexterity, machinery, or fieldwork?

Relationship & trust dependency10%

Does it rely on persuasion, emotion, conflict resolution, or trust?

raw =
(0.15·digitality + 0.20·language + 0.15·routine + 0.20·aiCapability + 0.10·adoption)
− (0.10·judgment + 0.10·physical + 0.10·relationship)
score = clamp( ((raw + 30) / 110) × 100, 0, 100 )

The raw range (−30 to +80) is mapped onto 0–100. We also derive sub-scores — automation, augmentation, human-moat strength, junior-role pressure, seniority protection, and reskilling urgency — from the same factors.

020
Low
2140
Moderate
4160
Significant
6180
High
81100
Severe
Step 7

Confidence levels

High — an exact O*NET/ESCO match, many task statements, and several sources agreeing.

Medium — a partial title match or a related occupation used as a proxy.

Low — an emerging or non-standard title; the result is directional and should be read as such.

Step 8

Source hierarchy

Tier 1 — scoring
O*NET, ESCO, ILO, OECD, OpenAI/OpenResearch, Microsoft Research, Anthropic Economic Index

Official occupational datasets and institutional research. Used to set scores.

Tier 2 — context
BLS Occupational Outlook Handbook, national labour datasets

Government labour data for outlook, wages, and descriptions.

Tier 3 — framing
WEF, McKinsey, Brookings, NBER, academic papers

Reputable analysis used for framing, never as the sole basis for a score.

Tier 4 — context only
News & media

Used for colour and context — never for scoring.

We do not use random blogs, social posts, or influencer predictions as scoring sources.

Step 9

Limitations

  • AI capability and adoption change quickly; every score is a point-in-time estimate, not a forecast.
  • Adoption varies by country, industry, company size, regulation, cost, and trust.
  • A job title can mean very different work in different organisations.
  • High exposure does not equal unemployment — it indicates where AI can assist or accelerate.
  • This tool is guidance, not career, legal, or financial advice.
Step 10

Update policy

We review scores as major occupational datasets refresh and as new AI labour-market research is published. Each result page shows when it was last updated. When the underlying evidence changes enough to move a score or a recommendation, we revise the page and the date.

See the method applied to a real role →