by Claude Opus 5.5

Which skills and capabilities are most likely to remain valuable or become more valuable as AI tools become widely available?

The skills that hold or gain value are the ones that decide what good looks like and carry the consequences. That means domain knowledge deep enough to catch a plausible error, the ability to specify and check work, operational know-how about how things actually get done, and the human and physical work around the output. Producing a competent first draft, whether a summary, a standard letter or boilerplate code, is losing value fastest, because that is now what the tools do.

What the 2026 UK evidence shows

Four pieces of 2026 evidence are worth knowing, because together they show employers asking for different things, not just fewer people.

  • Skills are changing faster where AI reaches. PwC’s 2026 UK AI Jobs Barometer (June 2026) finds that the skills required in the most AI-exposed jobs are changing about twice as fast as in the least exposed. The most exposed occupations picked up an average of 224 new skills, against 101 in the least exposed group. If you work in an exposed job, your skills have a shorter shelf life than they used to.

  • AI skills carry a premium, but read it carefully. PwC puts the UK wage premium for jobs requiring specialist AI skills at 34.2% in 2025, up from 11% in 2024. It ranges from 12% in government to 64% in consumer markets. These figures come from advertised salaries for roles that ask for those skills. They do not mean that finishing an online course will raise your pay by a third.

  • AI is spreading into ordinary job adverts. Indeed found that 9.4% of UK postings mentioned AI or AI tools at the end of June 2026, a record. The share was 48.8% in data and analytics and under 1% in cleaning, personal care and driving.

  • Employers want production-ready skills, not general ones. DSIT and LinkedIn’s June 2026 snapshot of entry-level hiring found 30 of 38 tracked entry-level roles shrinking, including accountants (−29%), graphic designers (−28%) and software engineers (−27%). Retail assistants (+25%) and sales development representatives (+17%) grew. The report’s most useful sentence for individuals describes a mismatch. Employers seek “specific, operational, production-ready capabilities (data governance, DevOps, VAT processing, consultative selling)”, while candidates offer “general, analytical skills (Python, SQL, legal research, financial accounting, CRM)”. The authors stress that the data are not causal evidence of AI’s impact.

PwC adds one more signal, from US data: AI-exposed junior roles are seven times more likely than the least exposed junior roles to ask for traditionally senior skills such as leadership and strategic thinking. Taken together, the pattern is that general analytical competence is being commoditised, while the ability to apply it inside a real process, and to own the result, is being bid up.

Seven capabilities likely to gain value

1. Domain judgement good enough to catch the subtle error. The best-known field experiment on this, with 758 Boston Consulting Group consultants using GPT-4 (Dell’Acqua et al., 2023), found that on tasks within the model’s capability they completed 12.2% more tasks, 25.1% faster and at much higher quality. On a task just outside that capability, AI users were 19 percentage points less likely to reach the correct answer than those without it. The authors called the line between the two a “jagged frontier”, and nobody can draw it in your field for you. Knowing where it runs is a skill only domain experts have.

2. Specifying and verifying work. As drafting gets cheap, the scarce work is writing a precise brief and defining what “correct” means. Examples include a checklist for reviewing an AI-summarised lease, test cases for an automated holiday-pay calculation, or acceptance criteria for code an assistant has written. People who can check output quickly and reliably become the bottleneck that every AI workflow needs.

3. Operational, process-specific know-how. The DSIT/LinkedIn list is telling: VAT processing, data governance, DevOps. These are skills of knowing how work actually moves through systems, controls and exceptions in a real organisation. A model can explain VAT in general. It cannot tell you why your firm’s partial-exemption calculation keeps going wrong.

4. Accountability that someone has to hold. Many UK roles carry named responsibility that cannot be delegated to software: Senior Managers and Certification Regime holders in finance, solicitors and qualified accountants, gas engineers, registered nurses. Since the Data (Use and Access) Act’s automated decision-making reforms commenced on 5 February 2026, organisations that make significant automated decisions about people must offer safeguards that include human intervention. That human has to be capable of disagreeing with the system. The ICO’s March 2026 review of AI in recruitment found “no meaningful human involvement” in many processes it examined, so the skill is in demand, not just required on paper.

5. Relationship, persuasion and care. David Deming’s research found that jobs requiring high social interaction grew by nearly 12 percentage points as a share of the US workforce between 1980 and 2012, while technical jobs without interpersonal demands declined. The 2026 UK entry-level data point the same way, with sales and customer-facing roles growing while information-processing roles shrink. Negotiating a claim settlement, calming a distressed patient and selling a complex service all get more valuable as the paperwork around them gets cheaper.

6. Physical and situated work. Electricians, care workers, plumbers and paramedics will use AI for quoting, scheduling, diagnostics and notes, but the core task happens in a specific place with specific hazards. Indeed’s under-1% AI-mention rates for hands-on occupations reflect that.

7. Workflow and AI-operations skill. This is where the measured premium sits. It covers building a small automation, setting up a retrieval tool over your team’s documents, and measuring whether output has improved. It also covers knowing which data may go into which tool. It does not require you to become an engineer, but you do need to be the person who makes the tool work reliably for your team.

What is losing value, and why you should still learn it

Generic first-draft production, routine information processing and recalling facts that are a search away are all losing market value. But these skills are also how people build the judgement in point 1. A trainee who never drafts a contract clause from scratch struggles to spot a bad one later. The answer is to learn foundations differently, and to verify more and compose less, not to skip them. Question 4.5 covers the habits that prevent deskilling.

A worked example: a motor claims handler

The tasks losing value are keying first notification of loss data, chasing documents and writing standard letters, all of which tools increasingly handle. The tasks gaining value are:

  • spotting fraud indicators that the model’s training data never saw;

  • recognising a vulnerable customer and adapting handling, as the FCA’s vulnerability guidance and the Consumer Duty expect;

  • negotiating liability on a disputed multi-vehicle claim;

  • explaining a decision to someone who is angry.

The handler who invests in the second list, and who learns to supervise the tools doing the first, is building towards complex claims, counter-fraud or vulnerability specialism. Question 4.7 covers those pivots in detail.

Bottom line

Aim to be the person who can say whether the output is right and who is answerable if it is not. That rests on deep domain knowledge, fast verification and a close understanding of how work flows through your organisation. Add enough AI-operations skill to make the tools useful, and keep building the human and physical capabilities that sit around the output. Because skills in exposed jobs are now changing twice as fast, expect to refresh this mix every year or two rather than once a decade.

Sources

From AI and Jobs: UK, October 2026