Which careers are most exposed to AI? We scored 1,324 of them

Most writing about AI and jobs falls into one of two modes: a list of ten roles that will supposedly be gone by 2030, or a reassurance that AI will “augment, not replace”. Neither is much use if you are trying to make an actual decision about your career.

So we did the work. We assessed all 1,324 careers in our library against five weighted dimensions of automation pressure, published the score on every career page, and are putting the full distribution here.

Some of what came out was uncomfortable, including for us. That is in here too.

What we measured, and what we deliberately didn’t

The score runs 0–100 and combines five dimensions, four of which carry weight:

• Core task exposure (35%) — how much of the role’s important work current AI could perform
• End-to-end automation (25%) — whether AI could complete whole workflows without a person in the loop
• Adoption pressure (20%) — how likely employers actually are to introduce it, accounting for cost, regulation and risk appetite
• Human dependence (20%, reverse-scored) — how much success depends on judgement, relationships and accountability
• Role adaptability — measured and shown, but deliberately excluded from the score

That last exclusion is a judgement call worth explaining. A role that can reshape itself is not under less automation pressure — it is better placed to absorb it. Folding adaptability into the score would quietly turn a pressure measure into a survival measure, and those are different questions.

One thing this score is not: a probability that a job disappears. It measures how much of the work may change. A role can score 80 and still employ more people in ten years, doing a different version of the job. We enforce that distinction in code — our build fails if any copy phrases the number as a likelihood of replacement — because it is the single easiest way for this kind of data to mislead.

The distribution

Across 1,324 careers:

Level	Careers	Share
Low (0–40)	508	38%
Moderate (41–60)	490	37%
Strong (61–100)	326	25%

The shape matters. A quarter of careers face strong automation pressure — high enough to take seriously, far from the “half of all jobs” figures that circulate.

It is also worth saying what this distribution replaced. Our previous automation ratings, generated one role at a time, came out 63% low and 2% high. That is what you get when you ask a model to assess a single job in isolation: it reassures. Every score here was produced by comparing roles against each other and against fifteen hand-scored reference points, precisely because the isolated version was not credible.

The most exposed roles

The top of the list is unambiguous and consistent:

Score	Career
93	Switchboard Operator
91	Data Entry Operator
90	Medical Transcriptionist
88	Payroll Clerk
87	Accounts Clerk
87	General Clerk
87	Telemarketer
86	Clinical Coder

Every one of these has the same structure: high-volume work on structured information, with a verifiable correct answer and no requirement for physical presence or personal accountability. Medical transcription is the clearest case — speech recognition has already taken most of it, and the remaining work is validating machine output rather than producing the transcript.

Notice these are not “unskilled” roles. Clinical coding requires real expertise. Exposure tracks the shape of the work, not its difficulty.

The least exposed roles

Score	Career
8	Footballer
10	Watch and Clock Repairer
11	Shearer
12	Neurosurgeon
13	General Surgeon
13	Hazardous Materials Removal Worker
13	Dancer
14	Singer

Two clusters, one obvious and one less so. The obvious one: physical work in unmapped environments, where the constraint is manipulation rather than cognition. The less obvious one: work where a named human carries the consequence. A surgeon’s exposure is low not because AI cannot read a scan — it can, well — but because someone has to take consent, make the call mid-operation, and answer for the outcome.

By family: where the pressure actually sits

Averaging across job families produces the finding most likely to surprise people.

Most exposed:

Avg	Family
65	Marketing & Brand
64	Journalism & Publishing
64	Architecture & Spatial Design
64	Accounting & Audit
63	Finance & Investment

Least exposed:

Avg	Family
40	Agriculture & Food
39	Social Services
38	Nursing & Midwifery
38	Mental Health
26	Skilled Trades

Skilled Trades — 122 careers — averages 26. Marketing & Brand averages 65. The decade of advice telling people to leave the trades for a desk has aged badly, and the gap is not close.

The reason is structural. Automation has historically moved from physical labour upward. This wave moved in the opposite direction: it started with language, images, code and analysis, which is precisely what the professional-class desk job is made of. Plumbing was never the target.

What we got wrong the first time

Two errors, both instructive, and both visible in our own data before we caught them.

We initially scored physical work as heavily automatable. Dishwasher came out at 89, Mushroom Picker at 89 — because the assessment credited robotic harvesters and dishwashing machines to AI. That is a category error. Those are capital equipment on a completely different timeline, and the dimension explicitly asks what AI could perform. Dishwasher is now 23.

Then we over-corrected. Having decided robotics was out of scope, we scored Rideshare Driver at 35. That is worse. Autonomous driving is AI — the vehicle was never the bottleneck, perception and judgement in traffic were, and driverless services now carry paying passengers commercially. Rideshare Driver is now 79.

The rule we landed on is to ask what the binding constraint actually is:

• If AI is the constraint and it is being solved, exposure is high — even for physical work. Rideshare Driver: 79.
• If physical manipulation is the constraint and AI is incidental, exposure is low. Vegetable Picker: 32. Machine vision already grades produce perfectly well; no gripper handles ripe produce without bruising it.

Both are physical jobs involving hardware, and they sit 47 points apart for a reason that can be stated.

Where a low score would still leave a false impression, the career page says so directly. Vegetable Picker’s assessment carries a note that robotic harvesters are improving quickly and the barrier is an active engineering problem rather than a permanent limit. A low AI score does not mean nothing is coming.

What actually protects a career

Reading across the whole set, four things separate the low scores from the high ones — and only one of them is about being clever.

Physical presence in an unmapped environment. Not physical work generally. Specifically work in places that are different every time.

Named accountability. Roles where a person signs, consents, certifies or is liable. The judgement can often be assisted; the responsibility does not transfer.

Relationship as the product. Where trust built over time is the service, not packaging around it.

Owning the standard rather than applying it. This is the one people miss. Executing a defined procedure is heavily exposed. Deciding what the procedure should be, and being accountable for it, is not. It is the difference between Accounts Clerk at 87 and a senior audit role in the fifties.

How to actually use this

Look up your own role. Every career page carries the full breakdown — the five dimensions, what AI may take on, where people remain essential, and how the role may evolve. The number on its own is the least interesting part.

Read the dimensions, not the score. Two careers can both score 60 for opposite reasons — one with high task exposure held down by strong human dependence, another moderate on everything. Those imply completely different responses.

Move toward judgement and accountability within your field. For most people the useful move is not a career change. It is moving up the stack from execution toward ownership, in the field they already know.

Treat it as a current view, not a forecast. We date every assessment and revisit it. AI capability is moving faster than any three-to-seven-year projection can honestly track, and anyone who tells you otherwise is selling something.

---

Methodology: 1,324 careers, five dimensions, scored in same-family batches against fifteen hand-scored calibration anchors, with the overall score computed arithmetically from the dimensions rather than assigned directly. Every assessment was then reviewed by an independent adversarial pass; roughly a fifth were flagged and regenerated. The full breakdown appears on each career profile.