by Claude Opus 5.5

Which new roles and capabilities typically emerge when organisations scale AI, and why?

Scaling AI creates fewer brand-new job titles than people expect and more new responsibilities inside existing jobs. The roles that do appear cluster around three gaps that open once AI moves from pilots into everyday operations: someone has to make it work reliably, someone has to own the risk, and someone has to change how the work is actually done. Agentic systems add a fourth gap: someone has to supervise software that takes actions rather than just drafting text.

Why new roles appear at all

The economic logic is complementary investment. Brynjolfsson, Rock and Syverson’s “productivity J-curve” (2021) argues that general-purpose technologies pay off only after firms invest in intangibles (redesigned processes, training, data, management routines) whose costs arrive first. New roles are the human form of those intangibles. A pilot can run on one enthusiast’s goodwill. A production system needs someone who is accountable when the output is wrong, when the vendor quietly updates the model, or when the Information Commission (which replaced the ICO on 30 September 2026) asks how a decision about a person was reached.

The UK hiring data show the specialist end of this clearly:

  • PwC’s 2026 AI Jobs Barometer (15 June 2026) found UK postings for specialist AI roles up 61% in 2025, to about 180,000, or 2.2% of all postings. The AI-skills wage premium rose to 34.2%, from 11% in 2024.

  • LinkedIn reported in June 2026 that about 95,000 AI roles had been created in the UK since 2023, even as overall hiring fell.

  • The Lloyds Business Barometer (July 2026) found 20% of firms creating AI-specific roles.

  • Indeed reported that 9.4% of UK postings mentioned AI at the end of June 2026, a record, rising to 48.8% in data and analytics.

The picture for the typical organisation is different, though. ONS data published in July 2026 show about 35% of businesses with 10 or more staff using AI, but the average user firm runs only 1.6 AI technologies, up from 1.4 in late 2023. About 62% of firms citing a lack of AI expertise are training or retraining staff, but only 11% have trained more than half their workforce. In most UK organisations the “new roles” are still part-time hats worn by existing staff.

The roles that recur

  • Platform and integration engineering. What it does: Connects models to data, identity and core systems; manages cost and vendor changes. Why it appears: Most value sits in integration, not the model. Usually grows out of: Software and data engineering.

  • Evaluation and quality. What it does: Builds test sets, measures accuracy, re-tests when models change. Why it appears: Generative output varies, and vendors update models without notice. Usually grows out of: QA, analysts, domain experts.

  • AI-enabled process owner. What it does: Owns the outcome metric for a workflow and redesigns it. Why it appears: Tools bolted onto unchanged processes rarely pay. Usually grows out of: Operations managers, business analysts.

  • Governance, model risk and assurance. What it does: Keeps the inventory, tiers risk, approves deployments, audits. Why it appears: Accountability needs a named owner. Usually grows out of: Compliance, data protection, internal audit.

  • AI security. What it does: Defends against prompt injection and data leakage; sets agent permissions. Why it appears: AI widens the attack surface. Usually grows out of: Information security.

  • Enablement and change. What it does: Runs training, champions networks and playbooks. Why it appears: Usage without skill produces rework. Usually grows out of: HR, learning and development.

  • Agent supervision. What it does: Watches queues of agent actions, handles exceptions, tunes permissions. Why it appears: Software now acts, so someone must oversee the actions. Usually grows out of: Operations team leaders.

Three of these deserve more comment, because they are the ones organisations most often under-resource.

Governance and assurance. Financial services is the leading indicator. In the Bank of England and FCA survey published in November 2024, 84% of firms said they had an accountable person for their AI framework, yet 46% reported only “partial understanding” of the AI technologies they use, largely because a third of use cases were third-party implementations. Being accountable for something you only partly understand is exactly the condition that creates demand for model-risk and assurance staff. The PRA’s model-risk expectations (SS1/23) already apply, and the industry is openly asking how they scale to generative and agentic systems. A fourth BoE/FCA survey closed on 31 July 2026, with results due by the end of the year. One caveat: no reliable UK data on hiring for compliance or model-risk roles specifically could be found, so the growth here is visible in firms’ structures more than in labour-market statistics.

Outside finance, data protection is the main pull. The Data (Use and Access) Act’s new automated decision-making rules commenced on 5 February 2026. They permit more automated decisions but require safeguards: information, a way to contest, and human intervention. The ICO’s March 2026 “Recruitment rewired” review found many recruitment tools operating with “no meaningful human involvement”. Designing human review that is genuinely meaningful is a skill, and it is landing on data protection officers and HR operations teams.

Evaluation. This is the most underrated capability. A team that holds a few hundred real, anonymised test cases with agreed correct answers can re-run them overnight when a model or prompt changes. A team without one finds out from complaints. Evaluators are rarely data scientists: the best test sets are written by the people who know what a good claims decision, contract summary or customer reply looks like. Many organisations make the mistake of hiring the engineers and forgetting to give domain experts the time to do this work.

Agent supervision. This is real but still small. The UK government’s AI adoption research (fieldwork in 2025) found only 7% of adopting firms using agentic AI, and in the 2024 BoE/FCA survey only 2% of financial-services use cases involved fully autonomous decision-making. The capabilities it needs, though, are distinctive: identities and permissions for non-human actors, logs of every action, and exception queues staffed by people who can tell a reasonable agent decision from a bad one. Question 3.13 covers this in detail.

What changes inside existing jobs

The larger shift is diffuse. PwC’s analysis of 2.4 million US entry-level roles finds that those most exposed to AI are now seven times more likely to require traditionally senior skills such as leadership and team-building. These “seniorised” US roles have grown 35% since 2019, while other US entry-level roles have fallen by 10%. The new capability is judgement and supervision pushed further down the hierarchy.

The Warwick projections published by DSIT in January 2026 point the same way. They expect jobs involving AI activities to grow from about 158,000 in 2024 to about 3.9 million by 2035, roughly 12% of the workforce. Most of these will be existing occupations with AI tasks added, not AI specialists. Managers increasingly decide what to delegate to software and how to check it. Experienced professionals become reviewers and test-writers. Team leaders absorb exception handling.

Sequencing: build, borrow or grow

A pattern that works for mid-sized UK organisations:

  1. Start with a named business owner for each production use case and a small central team covering platform, governance and security. Don’t appoint a chief AI officer before you have anything in production for them to run.

  2. Borrow evaluation and security expertise from vendors or advisers early, but bring it in-house once a core process depends on AI. Nobody else will notice when your system degrades.

  3. Grow rather than hire where you can. A 34% wage premium makes external AI hiring expensive. Retraining analysts as evaluators, compliance staff as AI assurers and team leaders as agent supervisors is cheaper, and it keeps the domain knowledge that the roles depend on.

What to watch

The BoE/FCA survey results, due by the end of 2026, will show whether governance in finance is keeping pace with adoption. A narrowing AI wage premium would signal that AI skills are diffusing into ordinary jobs.

Sources

From AI and Jobs: UK, October 2026