How the AI-proof score works

Six ratings, one formula

We rate each of 163 occupations from 0 to 5 on six factors. Four of them protect a job from automation and two expose it. The score is then calculated from those ratings with the same formula for every job, so nothing is hand-picked.

• ProtectsHands-on — skilled physical work in the real world.
• ProtectsUnpredictable settings — every site, home or patient is different.
• ProtectsHuman touch — trust, care or persuasion face to face.
• ProtectsAccountability — licensed or high-stakes — someone has to sign it off.
• ExposesScreen-based — work done on a computer with text, data or images.
• ExposesRoutine — repeatable, rules-based tasks.

Protective factors compound: one strong bottleneck is enough to make a job hard to automate. A plasterer scores 84 because the work is skilled, physical and different on every site, even though it isn’t especially people-facing. Being hands-on only counts fully when the setting is unpredictable, because robots already handle physical work on tidy, repetitive production lines. Screen-based and routine work then pulls the score down, because that’s exactly what AI is good at.

hands      = hands × (0.5 + 0.5 × variety)
protection = 1 − (1 − 0.9·hands)(1 − 0.45·variety)(1 − 0.7·human)(1 − 0.5·stakes)
exposure   = 0.6·screen + 0.4·routine
score      = 100 × (0.1 + 0.9·protection) × (1 − 0.5·exposure)
(each rating scaled 0–1)

What the bands mean

• Very safeAI won't replace this any time soon.
• Fairly safeAI will help with parts of the job, but it needs a human.
• ChangingAI will reshape this job, and fewer people may be needed.
• At riskAI can already do much of this work.
• High riskMuch of this work can already be automated.

What it’s based on

We don’t copy any single study’s numbers. Instead we rate each occupation using the factors these studies found matter most, and apply the same formula to all of them.

• The Future of Employment — Frey & Osborne, University of Oxford (2013)

  Identified the 'bottlenecks' that make jobs hard to automate: dexterity in messy environments, social intelligence and creative judgement.

• The probability of automation in England — Office for National Statistics (2019)

  Applied automation risk to the English workforce and found routine, lower-skilled roles most exposed.

• The impact of AI on UK jobs and training — Department for Education (2023)

  Mapped exposure to modern AI across UK occupations. Office, finance and professional jobs are the most exposed.

• GPTs are GPTs — Eloundou, Manning, Mishkin & Rock (2023)

  Measured how many tasks in each job large language models could speed up. Screen-based, text-heavy work scored highest.

• Anthropic Economic Index — Anthropic (2025 onwards)

  Tracks which work tasks people actually use AI for in practice, which is concentrated in software, writing and analysis.

What it can’t tell you

• Exposed isn’t the same as replaced. AI often changes a job rather than removing it. A low score means much of the work could be automated, and fewer people may be needed to do it.
• It’s about the job, not you. Your employer, your specialism and how you use AI all matter. People who learn to use AI well tend to do better in exposed jobs.
• Things change. If robotics speeds up, physical jobs will become more exposed. We’ll update the ratings as the evidence moves.

Where the jobs come from

Listings come from the official Reed API. We refresh them every few hours and remove ads after 30 days. Each job title is matched to an occupation by keyword rules, with AI used to classify unusual titles. You always apply on the original site.