AI in the NHS and the Public Sector

by Claude Opus 5.5

In UK healthcare and the public sector (including the civil service’s own AI deployment), what are the biggest practical blockers to scaling AI, and what workforce impacts follow from those blockers?

The tools are no longer the main constraint. Ambient voice technology, Copilot and the civil service’s own “Humphrey” tools are all in use. What stops them spreading is the system around them: legacy IT, information governance, fragmented procurement, clinical safety sign-off, weak evaluation and cautious industrial relations. Those blockers have three workforce effects. Time savings are banked as capacity rather than as fewer posts. A new layer of assurance and integration work appears that the public sector struggles to pay for. And headcount cuts that are happening, such as abolishing NHS England, come from restructuring decisions, not from AI.

Where deployment has got to

In the NHS, the leading use is ambient voice technology (AVT): AI scribes that listen to a consultation and draft the notes and letters. The 10 Year Health Plan (July 2025) committed to a framework procurement in 2026–27, citing local trials in which AVT saved one to two minutes per GP appointment. It estimated that saving 90 seconds per appointment would release “over 2,000 full time equivalent worth of GP capacity”. Regions have moved faster. NHS England Midlands, with Dudley and Sandwell and West Birmingham trusts leading, has contracted Heidi to cover about 70,000 clinicians across 15 trusts and 1,239 GP practices, with go-lives from May 2026. AI also supports about a third of NHS chest X-rays, some 2.4m scans.

In central government, the “Humphrey” suite launched in January 2025. It includes:

Consult, which analyses consultation responses. It processed more than 50,000 responses to the Cunliffe water review in two hours for £240, and DSIT said it could free 75,000 staff days a year.
Minute, a meeting scribe used by 1,000 officials across 22 councils.
Parlex, Redbox and Lex, for parliamentary research, document summaries and legal research.

HMRC began rolling out Microsoft 365 Copilot in October 2025 with 32,000 licences, planning to reach 50,000 in 2026. It is the largest government deployment of Copilot so far.

Headcount: AI is not what is moving the numbers

NHS. Employment was 2.07m in June 2026, down 1,000 on the year, which is flat.
Civil service. Headcount was 557,000 in June 2026, down 1,000 on the quarter (the first quarterly fall since June 2021) but up 6,000 on the year. Full-time-equivalent staffing reached 524,000 in March 2026, the highest since 2005.
Cabinet Office. Staffing fell 13% in a single quarter (September–December 2025) through a voluntary exit scheme that was part of a planned one-third reduction. The department has said its planned savings will come partly from technological efficiencies.
HMRC. Even while rolling out Copilot, HMRC has kept growing. It added about 1,100 staff in the figures to March 2026 as it hires compliance and debt-management staff funded at the Spending Review.
Targets. The government aims to cut back-office spending by 16% by 2030 but has set no numeric headcount target and no AI-specific one.

The biggest cuts are organisational. The NHS Modernisation Bill, announced in the May 2026 King’s Speech and at Public Bill Committee stage in July, abolishes NHS England and moves its functions into the Department of Health and Social Care. About 18,000 administrative posts across NHS England and DHSC were reported in November 2025 to be going, with about £1bn approved for redundancy costs, and integrated care boards are being cut by about half. Those cuts would happen with or without AI. They also remove some of the people who would run national AI procurement and evaluation.

The blockers, and the workforce impact of each

Legacy IT. The government’s State of Digital Government Review (January 2025) estimated that 28% of central government systems were legacy in 2024, up from 26% in 2023. Among NHS trusts the share ranged from 10% to 70%. NHS England’s AVT guidance calls integration with the electronic patient record “essential”. In a trust with an old patient record system, a scribe that cannot write back into the record just adds a copy-and-paste step. When the plan was published, the chair of the Royal College of GPs said many GPs report basic IT systems that are “slow, inefficient and can’t communicate with one another effectively”. Workforce effect: savings are uneven. Clinicians in digitally mature trusts gain time, while others gain extra clicks, so the national productivity figures understate what works locally and overstate what scales.

Information governance. Each deployment needs a data protection impact assessment. NHS England’s guidance says the NHS organisation keeps a non-delegable duty of care whatever its contract with the supplier says, and NHS organisations, GP practices included, have been warned they “may still be liable” for negligence claims arising from AI use. Workforce effect: the demand for information governance leads, data protection officers and Caldicott Guardians rises with every deployment. In a stretched trust, these few specialists become the bottleneck.

Procurement. NHS England’s AVT Supplier Registry is a self-certified list. Its guidance states: “This is not a commercial framework. Procurement will be carried out by individual NHS bodies.” So every trust or region runs its own process until the national framework arrives. The Midlands deal shows that scale is possible when a region pools its efforts, but most have not done so yet. Workforce effect: commercial and digital teams duplicate effort, and their numbers are being cut in the ICB and NHS England reductions.

Clinical safety. Under the NHS clinical risk standards, the supplier must hold a DCB0129 safety case and the deploying organisation must complete DCB0160 documentation with its own hazard log. Products also need at least Class I medical-device registration with the MHRA, and AVT that summarises using generative AI is likely to count as a higher-function device. Workforce effect: every deployment needs a trained clinical safety officer, usually a clinician doing it alongside their main job. This is the clearest case where AI creates skilled work before it saves any.

Evaluation. Many of the headline numbers are projections. The £45bn a year cited at Humphrey’s launch is the review’s estimate of the gain from full digitisation of public services, not from AI alone. Consult’s main test showed it agreeing with one or both expert panels about 83% of the time, and the two expert panels agreed with each other only 55% of the time. That is encouraging but comes from a single exercise. The 2,000-FTE GP figure is a government estimate extrapolated from local trials, not a measured result. The Midlands pilots report an 80% cut in documentation time in same-day emergency care, a trust claim that has not been independently evaluated. Workforce effect: without solid evaluation, finance directors cannot bank savings. So time released goes into extra appointments and shorter backlogs (which is often the stated aim), and staffing budgets stay the same.

Union relations and staff confidence. HMRC reports that some staff objected to Copilot or worried about misuse. GP reaction to the 90-second estimate in the trade press has included outright dismissal and doubt that saved time would become extra appointments. The Employment Rights Act’s new union access rights start on 30 October 2026. The TUC’s “pro-worker AI” strategy calls for mandatory impact assessments and human review of high-risk decisions. No major 2026 public-sector dispute specifically about AI was found. But the private-sector warning sign is clear: when British Gas cut 1,300 call-centre roles, unions rejected the company’s claim that AI was not the cause. Workforce effect: public employers are framing AI as freeing up time, not cutting jobs, partly to keep staff on side, and that framing limits how far savings can be turned into lower headcount.

The wider workforce consequences

Three effects follow from these blockers together.

Capacity, not cuts. The workforce effect of AI in health and government in 2026 is extra activity within the same headcount. In politics and in clinical practice, that is easier to defend than redundancies.
A thin layer of expertise. The roles that unlock scaling (integration engineers, clinical safety officers, information governance leads, product-minded commercial staff) are scarce and poorly paid relative to the private sector. The digital government review found central government cyber specialists earning about 35% less than private-sector peers, and only 50% of digital recruitment campaigns succeeding in 2024, down from 80% in 2020.
Restructuring before redesign. The administrative cuts in NHS England and the ICBs are happening before the tools that might absorb that work have been evaluated at scale. If the cuts land first, remaining staff will carry the work while the tools are still being bought and assured.

What to watch

The national AVT framework procurement through 2026–27.
The NHS Modernisation Bill’s progress and the timing of the transfer from NHS England to DHSC.
Whether the Cabinet Office, now home to AI strategy after DSIT’s abolition, publishes any evaluation of the Humphrey tools.
HMRC’s results as it reaches 50,000 Copilot users.
Public-sector employment figures, which have so far shown no sign of AI-driven change.

Sources

From AI and Jobs: UK, October 2026