by Claude Opus 5.5

Which job families are most likely to shift toward AI-supervised exception handling versus AI-assisted production versus partial displacement, and what drives those differences?

High-volume, rules-based processing work is moving towards AI-supervised exception handling. Examples include insurance claims, accounts payable and much of conveyancing. In these jobs software handles the routine flow and people deal with what it cannot or should not decide. Work whose output is a crafted artefact with an accountable author is moving towards AI-assisted production: code, marketing content, lesson plans and clinical notes. Partial displacement is most likely where volumes are high, the work is already digital, and cheaper output does not create more demand. Front-line customer service and transactional finance fit that description. Hands-on, relational work such as skilled trades and care is the least affected.

Three trajectories, defined

  • AI-supervised exception handling. The system processes most cases end to end. Humans set the rules, monitor quality and take the hard cases: disputes, fraud flags, vulnerable customers, and anything that falls outside policy. Teams get smaller and more senior, and the work gets harder per case.

  • AI-assisted production. A person still authors the output and is accountable for it, but drafts, code, research and formatting are accelerated. Headcount depends on whether demand grows as output gets cheaper.

  • Partial displacement. A meaningful share of the work disappears, not just the tasks within it. It usually happens through attrition and hiring freezes rather than mass redundancy, and the remaining roles shift towards exception handling.

Most job families show all three to some degree. The question is which one dominates.

How much is measured, and how much is judgement

The list below is a judgement based on task structure and early evidence, not a forecast. Some of the evidence is already measured.

  • Vacancies in exposed occupations. Bank of England staff analysis on the Bank Underground blog (August 2026) found that vacancies in the most AI-exposed occupations fell 15% over three years, against 6% in the least exposed. Customer service vacancies fell 23% and admin 22%. It is staff research, not Bank policy, and it says “confident attribution remains premature”.

  • Entry-level hiring. DSIT and LinkedIn data show entry-level hiring falling in information-processing roles. The authors caution that this is “not causal evidence”.

  • The wider economy. Cost pressures (employer NICs, the National Living Wage) and weak demand explain much of the hiring slowdown across the whole economy.

Treat the trajectories as directions of travel, not outcomes already achieved.

The eleven families

  • Customer service (contact centres). Likely dominant trajectory: Partial displacement of routine contacts. Remaining staff move to exception handling: complaints, vulnerable customers, escalations. Main drivers: Very high volumes; conversations already digital and easy to measure; Consumer Duty and FCA vulnerability guidance require skilled human handling of harder cases. UK signal to date: BoE staff analysis: customer service and admin vacancies down more than 20%. British Gas is cutting 1,300 call-centre roles; the CEO says AI is not the driver and cites a 20% fall in call volumes, which unions dispute.

  • Insurance claims and underwriting. Likely dominant trajectory: Exception handling. Main drivers: Structured data; straight-through processing for simple claims and standard personal-lines risks; fraud, large losses, disputed liability and specialty risk still need judgement and negotiation. UK signal to date: Little published UK headcount data; direction inferred from task structure.

  • Accounts payable and receivable. Likely dominant trajectory: Partial displacement plus exception handling. Main drivers: Invoice capture, matching and cash allocation are highly standardised; demand for processing does not rise when it gets cheaper; humans keep mismatches, supplier disputes and credit-control conversations. UK signal to date: Entry-level accountant hiring −29% (DSIT/LinkedIn, Apr 2026); not attributed to AI by the authors.

  • Paralegal and conveyancing. Likely dominant trajectory: Conveyancing: exception handling. Paralegal: AI-assisted production. Main drivers: Search summaries, document review, ID checks and first drafts automate well; title defects, client care, fraud risk and regulated sign-off do not. UK signal to date: Entry-level legal assistant hiring −14% (DSIT/LinkedIn). Training contracts −1.7% across 100+ firms, which Legal Cheek attributes mainly to solicitor apprenticeships, with AI a “potential” factor.

  • Software development. Likely dominant trajectory: AI-assisted production, with an entry-level squeeze. Main drivers: Coding assistants raise output; demand for software has historically grown as it gets cheaper; review, architecture and security become the scarce work. UK signal to date: Entry-level software engineer hiring −27% (DSIT/LinkedIn). Graduate tech jobs −46% in 2025 (ISE, reported). Average advertised IT salaries +16.8% y/y (Adzuna, Jul 2026, reported).

  • Marketing content. Likely dominant trajectory: AI-assisted production, with partial displacement at the commodity end. Main drivers: Product descriptions, ad variants and social copy are now near-free; brand judgement, strategy, rights clearance and performance analysis are not. UK signal to date: Entry-level graphic designer hiring −28% (DSIT/LinkedIn). Creative industries employment 2.464m in 2025, no significant change on 2024 (DCMS).

  • HR advisory. Likely dominant trajectory: Exception handling. Main drivers: Policy questions and letter drafting automate well; employee relations cases, investigations and new employment law obligations rise; AI in recruitment brings data-protection duties. UK signal to date: CIPD: two-thirds of employers expect more admin from the Employment Rights Act’s zero-hours reforms. ICO found weak human oversight in recruitment automation (Mar 2026).

  • Clinical admin. Likely dominant trajectory: AI-assisted production, with partial displacement of transcription and typing. Main drivers: Ambient voice technology drafts notes and letters; clinicians remain accountable; coding, referrals and patient contact still need people. UK signal to date: A government estimate, not a trial result, puts AVT’s saving at about 90 seconds per appointment (reported). The Midlands procurement covers 70,000 clinicians. Separate NHS admin cuts stem from abolishing NHS England, not from AI.

  • Teaching. Likely dominant trajectory: AI-assisted production. Main drivers: Planning and resource preparation speed up; the classroom, relationships, behaviour and safeguarding do not; demand is set by pupil numbers and policy. UK signal to date: EEF trial: ChatGPT with a guide cut KS3 science planning time by 31%, with no noticeable quality difference (Dec 2024; quality finding to be “treated with caution”).

  • Skilled trades. Likely dominant trajectory: Largely unaffected in core work; AI-assisted admin. Main drivers: Physical, varied sites; liability and certification; AI helps with quoting, scheduling, diagnostics and compliance paperwork. UK signal to date: AI use in construction is 13% of firms with 10+ staff, against 58% in information and communication (ONS, Jul 2026).

  • Care. Likely dominant trajectory: Largely unaffected in core work; AI-assisted admin. Main drivers: Personal care is physical and relational; labour is scarce, so tools are used to stretch staff, for example in care planning, rostering and monitoring. UK signal to date: Adult social care vacancy rate 6.2%, about 96,000 posts, a ten-year low (Skills for Care, 2025/26, reported).

What drives the differences

Five factors explain most of the pattern.

1. How structured and digital the inputs are. Claims forms, invoices and chat transcripts arrive as data that a system can process end to end. A leaking boiler or a frightened care-home resident does not.

2. The cost of an error, and who carries it. Where a mistake has legal, financial or clinical consequences, a named human usually stays in the loop. Examples are a wrongly declined claim, a defective title or a misfiled medication change. Regulation reinforces this. The Data (Use and Access) Act’s automated decision-making safeguards, in force since February 2026, require a route to human intervention in significant decisions. The FCA’s Consumer Duty and vulnerability guidance expect staff who can recognise and respond to vulnerability. These rules tend to push work towards exception handling rather than full automation.

3. Whether demand expands when costs fall. Cheaper software has historically meant more software, which is the Jevons-style effect, so development may see output rise faster than headcount falls. Invoices do not multiply because processing them got cheaper, and nobody wants more calls to their energy supplier. Where demand is fixed, productivity gains turn into fewer jobs.

4. Physical presence and relationships. Work that is embodied, varied and built on trust is the hardest to automate and the slowest to change. Indeed’s data show AI mentioned in under 1% of postings for cleaning, personal care and driving, against 48.8% in data and analytics (June 2026).

5. Labour economics. In shortage sectors such as care, nursing and parts of teaching, AI is used to stretch scarce staff rather than replace them. In sectors facing cost pressure, automation is a substitute for hiring. The Bank of England’s Agents reported in February 2026 that firms planned to meet demand “by investing in automation and AI, rather than by raising headcount”. Its July 2026 Monetary Policy Report said firms are “often slowing hiring or leaving vacancies unfilled” in highly automatable jobs.

The entry-level catch

Exception handling needs experience. A complex claim, a fraud flag or a difficult employee relations case calls for judgement that people used to build by processing hundreds of routine cases. If the routine work goes to software, the training ground goes with it. This is why the sharpest falls in the 2026 data are at entry level: DSIT/LinkedIn found 30 of 38 entry-level roles shrinking. It also explains Morgan Stanley’s report that cuts are concentrated in roles needing two to five years’ experience. For individuals in these families, the risk is less that the job disappears than that the bottom rungs of the ladder do.

What this means for you

Judge which trajectory your daily tasks are on, not your job title. If most of your time goes on the routine flow, move towards the exceptions: the escalation queue, the complaints panel, the fraud referrals. In AI-assisted production, aim to become the reviewer who sets standards, not just a faster drafter. In trades or care, AI is mainly an admin tool, and learning it well is a modest but real advantage.

What to watch

  • Whether further Bank of England work firms up its staff’s finding that the most AI-exposed occupations are losing vacancies fastest.

  • The ISE Student Recruitment Survey, due on 14 October 2026.

  • Whether firms attribute further contact-centre cuts to AI or to falling volumes.

Sources

From AI and Jobs: UK, October 2026