Analysis · 4 November 2026 · 5 min read
What skills can't AI replace?
Job ads, Anthropic's own hiring and our index point to judgment, leadership and hands on work. What the evidence supports, and where it runs out.
No study can prove a skill is safe from AI for good. What the data can show is which skills employers ask for more as AI spreads, and which kinds of work it has barely reached. Both point the same way: judgment, leadership and dealing with people, plus hands on work in places that are never quite the same twice.
Which skills are employers asking for more?
One very large dataset is PwC's 2026 AI Jobs Barometer, which covers more than a billion job ads in 27 countries and territories. It splits roles into two tracks. In "professionalised" roles such as radiologists and recruiters, AI handles routine tasks and human expertise matters more. In "democratised" roles such as IT service managers and medical secretaries, AI makes the job easier for non experts to do. The first group saw twice the job growth and 42 percent faster salary growth.
The clearest result is at entry level. Across 2.4 million US entry level jobs, the roles most exposed to AI were seven times more likely to ask for skills such as leadership, creativity or face to face work. Openings for those roles grew 35 percent since 2019, while other entry level roles shrank 10 percent.
Two limits apply. Job ads show what employers ask for, not whether having the skill keeps anyone employed. And PwC is a consultancy that advises companies on AI.
Anthropic's report on its own workforce points the same way from a very different source. On its interview scorecards, the biggest increases for technical candidates were in comments on autonomy and judgment, while discussion of domain expertise shrank. For non technical candidates, customer and mission orientation rose and discussion of execution skills fell. That is one unusual company, but it is further along with AI than most.
Which kinds of work has AI barely reached?
Our index gives a second view, because it scores occupations rather than skills. The bottom of it is dominated by physical work with people or in unpredictable places:
Plumbers, Pipefitters, and Steamfitters (14), Electricians (17), Physical Therapists (22) and Registered Nurses (27).
Compare those with work done entirely on a screen by following set steps:
Data Entry Keyers (63) and Customer Service Representatives (68).
Most of the research behind these scores measures language and software AI, so a low score for physical work does not rule out robots. Our post on AI and robots looks at that separately, and our list of safer jobs goes through the occupations.
Do workers have this right?
Partly. In a 2026 Campbellsville University survey of about 1,000 US adults, workers ranked specialised trade skills as the most protected, at just over half, and emotional intelligence second at 45 percent. Only 7 percent named technical IT strategy. Seventy percent said ethical decisions should stay with humans above all other duties.
The trades answer fits our index, and the emotional intelligence answer fits the job ads. But these are opinions rather than measurements, and the university sells the programmes it promotes alongside the results.
Which "safe" skills are shakier than they sound?
A skill on the list does not make a job safe, because a job is a bundle of tasks.
Creativity is the obvious case. PwC names it among the skills employers want, yet Graphic Designers (47) sit in the middle of our index and Market Research Analysts and Marketing Specialists (64) sit high. Both need ideas, and both also involve a lot of screen work that AI handles well.
Domain expertise is another. It still matters, but it came up less on Anthropic's scorecards. Knowing a field seems to count for more when you use that knowledge to decide and check, and less when the job is mainly to retrieve it.
Writing is a third. One Anthropic product writing team saw much of its work handed to agents and moved into writing prompts and evaluations instead. The skill survived, in a different job.
How do you apply this to your own job?
Ignore the skill names and look at your week. Count the hours you spend deciding what to do, handling people who disagree, or dealing with situations that do not fit the procedure. Then count the hours spent producing something from clear instructions. The second group is what AI reaches first, whatever your job title says.
Our index scores 961 occupations from published research. The assessment adds your own answers on routine, relationships, creativity and specialisation to your job's score. It takes about two minutes, free, no account.
Where does your job sit?
Every number above is an occupation average. Your own exposure depends on your seniority, your sector and how much of your day is routine. Answer a few questions and get your personal score free.
Get my personal scoreFree · No credit card · No signup
Figures come from the AI Job Risk Index, the data behind our AI job risk calculator, and were current when this was published. Scores change as new research is ingested, so the index is always the live version. See how scoring works. Informational guidance based on published research, not professional career or financial advice.