by Claude Opus 5.5

What are the most credible UK-specific sources for tracking AI-and-jobs trends (ONS, Bank of England, OBR, regulators, industry bodies, consultancies, academic work), what does each do well or poorly, and which new sources proved useful in 2026?

There is no single best source. The ONS tells you what is happening to jobs but not why, the Bank of England’s surveys come closest to linking AI to firms’ hiring decisions, and job-board and consultancy data are fast but hard to verify. The sources that proved most useful in 2026 were the Bank’s Decision Maker Panel (DMP) AI questions, the ONS’s July synthesis of its AI survey data, a Bank staff blog linking AI exposure to vacancies, and the DSIT/LinkedIn entry-level snapshot.

Three questions, three kinds of source

Tracking AI and jobs means answering three separate questions:

  • Adoption. Are firms and workers actually using AI?

  • Demand. Are employers advertising for different roles and skills?

  • Outcomes. Are employment, hours, pay and productivity changing?

No UK source answers all three. Sources also differ in what kind of evidence they offer: measured counts (payrolls, job adverts), survey responses about the present (adoption), and expectations (what managers think will happen). In 2026 the gap between the last two was large. Firms expect real headcount effects, but nearly 90% of DMP respondents say AI has had no material effect on their employment so far (BoE, July 2026).

The main sources compared

  • ONS Labour Force Survey (LFS). What it measures: Employment, unemployment and inactivity by age, occupation and region. Strengths: The only household-based source that captures occupation, self-employment and young people outside payrolls. Weaknesses: Low response; ONS warns of volatility, especially for 2023–24. The transformed LFS is now “most likely” in November 2027, and ONS will not seek reaccreditation. Frequency: Monthly (rolling three months).

  • ONS payrolls (HMRC RTI). What it measures: Payrolled employees by age, industry and region. Strengths: Administrative data, large and timely. Weaknesses: No occupation, so it can’t show AI-exposed roles; excludes the self-employed; flash estimates get revised. Frequency: Monthly.

  • ONS BICS. What it measures: Firm AI use, which technologies, workforce effects. Strengths: Large, official, repeated, broken down by size and industry. Weaknesses: AI questions only in some waves (none in Wave 164, 24 September 2026); excludes finance and the public sector; self-reported. Frequency: Fortnightly survey; AI module periodic.

  • ONS “AI in UK businesses: 2023 to 2026” (20 Jul 2026). What it measures: Consistent adoption trend for firms with 10+ staff; headcount and training responses. Strengths: First consistent multi-year series; shows adoption is broad but shallow. Weaknesses: Uses a different base from headline BICS, which confuses readers; one-off article. Frequency: One-off.

  • BoE Decision Maker Panel. What it measures: CFOs’ expected and realised effects of AI on employment and productivity. Strengths: About 2,000 firms, rigorous methods, asks about both expected and realised effects. Weaknesses: Covers expectations more than outcomes; CFOs may miss frontline task change. Frequency: Monthly; AI questions periodic.

  • BoE Agents. What it measures: Qualitative business intelligence. Strengths: Picks up mechanisms early, such as slower replacement hiring and changes to role design. Weaknesses: Not quantified; depends on which contacts the Agents see. Frequency: Quarterly summary.

  • BoE Monetary Policy Report and Bank Underground. What it measures: MPC judgements; staff analysis. Strengths: Brings together the DMP, ONS and vacancy data; the blog is explicit about causality. Weaknesses: Bank Underground posts do not represent the Bank’s view. Frequency: MPR quarterly; blog ad hoc.

  • OBR (EFO, FRS). What it measures: Fiscal and productivity scenarios involving AI. Strengths: Puts numbers on AI scenarios, including tax-base effects. Weaknesses: Scenarios, not forecasts; AI is not quantified separately in the central case. Frequency: Twice yearly (EFO), annually (FRS).

  • DSIT/LinkedIn entry-level snapshot. What it measures: Year-on-year hiring for 38 entry-level roles. Strengths: Occupation-level detail that official data lack. Weaknesses: LinkedIn users are not representative; no causal claim; continuity uncertain after DSIT’s abolition. Frequency: One-off (June 2026).

  • DSIT AI adoption research. What it measures: Adoption, uses, barriers, agentic AI. Strengths: 3,500 firms; asks about human oversight and agentic AI. Weaknesses: Fieldwork from February–May 2025, so dated. Frequency: Occasional.

  • ICO. What it measures: How employers use automated decisions, especially in recruitment. Strengths: Regulator’s direct view of practice in 30+ employers. Weaknesses: Qualitative; no prevalence data. Frequency: Ad hoc.

  • FCA (with BoE). What it measures: AI use in financial services; AI Lab sandboxes. Strengths: Detailed sector data on use cases and autonomy. Weaknesses: Not about jobs; the 2026 survey’s results are not due until the end of the year. Frequency: Roughly every two years.

  • CIPD Labour Market Outlook. What it measures: Employers’ hiring intentions and expected AI headcount effects. Strengths: Around 2,000 employers; HR view; size and sector breakdowns. Weaknesses: AI headcount question last asked in autumn 2025. Frequency: Quarterly.

  • KPMG/REC Report on Jobs. What it measures: Recruiters’ view of placements, vacancies and pay. Strengths: Timely; good at spotting turning points. Weaknesses: No AI content in 2026. Frequency: Monthly.

  • Indeed Hiring Lab. What it measures: Postings, wages, share of adverts mentioning AI. Strengths: Near real time; transparent charts. Weaknesses: Postings are not hires; coverage varies by sector. Frequency: Continuous; regular notes.

  • Adzuna. What it measures: Vacancies, advertised salaries, a graduate category. Strengths: Long series; regional detail. Weaknesses: Graduate classification disputed. Frequency: Monthly.

  • Lightcast. What it measures: Postings coded to a detailed skills taxonomy. Strengths: Skills granularity; widely used by researchers and consultancies. Weaknesses: Proprietary and paywalled. Frequency: Continuous.

  • PwC AI Jobs Barometer. What it measures: AI-skill demand, wage premia, growth by exposure. Strengths: Very large dataset (more than a billion adverts globally). Weaknesses: Method partly opaque; headline-friendly framing. Frequency: Annual (June).

  • Morgan Stanley AlphaWise. What it measures: Self-reported net job effects in firms using AI. Strengths: Cross-country comparison. Weaknesses: Only AI-using firms; mainly reported via the press; sector coverage shifted between waves. Frequency: Occasional (Jan, May 2026).

  • Academic and think-tank work. What it measures: Causal designs, exposure indices, graduate outcomes. Strengths: Best on identification and method. Weaknesses: Slow; often US data. Frequency: Irregular.

What each does well and poorly in practice

Use the ONS to anchor the totals, not to attribute causes. In 2026 payrolls and the LFS told different stories: RTI payrolls were down 145,000 on the year by August, while LFS employees were up 111,000 (ONS, September 2026). Neither series can isolate AI by occupation. BICS covers firms rather than workers and leaves out finance, one of the most exposed sectors.

Read Bank of England evidence in layers. The DMP gives numbers: firms expect AI to cut employment by about 0.4% a year and raise productivity by about 0.9% a year over three years (July 2026 MPR). The Agents give the mechanism: AI is “influencing role design and replacement hiring”, with “limited evidence of broad AI-driven reductions in employment” (September 2026). Bank Underground posts are staff research, explicitly not Bank policy.

Regulators show practice, not counts. The ICO’s “Recruitment rewired” review (31 March 2026) found employers relying on solely automated decisions without meaningful human involvement, but it says nothing about how widespread that is. The FCA and BoE survey is the best sector source: in 2024, 75% of financial firms used AI but only 2% of use cases were fully autonomous.

Industry bodies are timely but went quiet on AI. CIPD’s finding that 17% of employers expect AI to cut headcount (autumn 2025) is still the latest from that series, because its spring and summer 2026 surveys did not repeat the question. KPMG/REC is a good guide to the cycle but does not mention AI.

Job-board data are fast but contested. Indeed’s count of adverts mentioning AI reached a record 9.4% in June 2026. Adzuna’s figure showing graduate vacancies down 45.6% was challenged by Prospects Luminate’s Charlie Ball, who estimates that Adzuna captures only 5–10% of graduate vacancies. Indeed’s comparable measure was down about 7%. Gaps this large are a reason to compare several boards rather than rely on one.

Consultancy and bank surveys get the headlines and should be read with the most caution. PwC’s barometer is useful for skills and wage premia (a 34.2% UK AI-skills premium in 2025). Morgan Stanley’s finding that the UK leads on net AI job losses (−8% in January, −6% in May, from waves covering different sectors) comes from a survey of AI-using firms, and the primary data were not published.

Academic work is slow but disciplined. Useful 2026 examples included Stanford’s “Canaries” update (US data) and Bloomberg Economics’ UK analysis (16 June), which argued that weak demand and costs explain job losses better than AI.

Which new sources proved useful in 2026

  1. The DMP AI questions. For the first time a representative UK firm survey put numbers on both expected and realised AI effects. The realised figure was close to zero.

  2. The ONS “AI in UK businesses: 2023 to 2026” article. It tracks adoption among firms with 10+ staff from about 12% in late 2023 to about 35% in 2026. It also shows depth barely changing: the average number of AI technologies per user rose only from 1.4 to 1.6.

  3. Bank Underground, “Canaries in the column?” (6 August 2026). It combined four exposure measures with ONS/Textkernel job-advert data and found adverts down 15% in high-exposure occupations, 10% in mid-exposure and 6% in low-exposure. It also concluded that “confident attribution remains premature”.

  4. DSIT/LinkedIn entry-level snapshot (8 June 2026). This was the first government-badged occupation-level look at entry hiring: accountants −29%, graphic designers −28%, software engineers −27%.

  5. OBR scenarios. The March 2026 EFO and the July 2026 FRS made AI a quantified fiscal risk for the first time.

  6. Deloitte’s survey of 25,000 workers (May–June 2026). It measured shadow AI at scale: 31% of GenAI users use it without their employer knowing.

What to watch

The most informative releases between now and spring 2027:

  • The next BICS wave with AI questions (date unknown).

  • Results of the BoE/FCA financial services survey, due at the end of 2026.

  • The ISE Student Recruitment Survey on 14 October.

  • Any repeat of the DSIT/LinkedIn snapshot now that AI policy sits in the Cabinet Office.

  • Reweighted LFS data, due towards the end of 2026.

When sources disagree, give the most weight to measured flows (payrolls, adverts by occupation), then to the DMP’s realised effects, and the least to firms’ expectations.

Sources

From AI and Jobs: UK, October 2026