Executive summary
A significant paradox is emerging in UK healthcare. Polling shows that NHS staff are overwhelmingly supportive of using artificial intelligence, with 76% backing its use for patient care and 81% for administrative tasks. This provides a clear mandate for innovation. Yet, real-world data reveals that UK clinicians report far lower actual use of AI than their peers in France, Germany, and Denmark. The primary barriers are a high fear of clinical errors (62%), low confidence in using the tools, and a lack of awareness of practical applications (Health.org.uk, Digital Health).
This gap between enthusiasm and adoption is a problem that needs solving. A compelling European counterpoint can be found in Spain, where the Madrid Health Service (SERMAS) has piloted DxGPT—a GPT-4–based tool for rare-disease differentials—in its primary care system. While not without its own challenges, the pilot offers valuable lessons. To close its own adoption gap, the UK must pair its strong national guardrails—like the MHRA AI Airlock and NHS England's guidance—with targeted training and use-case pilots that directly address the documented fears and confidence gaps of its clinical workforce.
The adoption paradox
Support is high…
The appetite for AI within the NHS is undeniable. A comprehensive July 2024 survey from the Health Foundation found that more than three-quarters of NHS staff (76%) support using AI to assist with patient care, and an even higher proportion (81%) back its use for administrative tasks. This represents a powerful political and organisational "permission" to act and innovate.
…but everyday use is not.
Despite this support, the UK is lagging in practical adoption. A pan-European survey of healthcare professionals conducted by YouGov/Corti in late 2024 revealed that 73% of UK HCPs had never used AI at work. This places the UK behind continental peers in France, Germany, and Denmark in terms of day-to-day use (Digital Health).
Why it matters.
This gap between support and use is more than a curiosity; it has direct implications for the NHS's core policy goals. Closing the UK's well-documented productivity gap, improving patient safety, and tackling health inequalities all depend on the successful and widespread adoption of effective digital tools.
Diagnosing the UK’s barriers
The YouGov/Corti survey provides a clear diagnosis of the UK’s specific barriers to adoption.
- Fear of clinical errors (62%): UK clinicians reported the highest levels of anxiety about AI-related mistakes compared to their European counterparts. This highlights a deep-seated concern around trust and safety, underscoring the need for "provenance-first" tools that show their working and provide clear citations.
- Lower confidence using AI: Fewer than one in four UK HCPs reported feeling comfortable operating AI tools, a figure below the European average. This is not a reflection on digital literacy, but a direct call for better, role-specific training.
- Awareness gap (35%): Over a third of UK HCPs stated they were simply unaware of practical, real-world AI use-cases, such as automated note-taking or clinical decision support. This points to a need for clearer communication and procurement signals from national and local leaders.
European case study: Spain’s DxGPT in public health
While the UK grapples with these barriers, other health systems are moving forward with pilots.
- What it is: DxGPT is an AI assistant built on GPT-4, designed to take a clinical vignette and return a ranked, reasoned top-five differential diagnosis, with a focus on rare and complex cases.
- How Spain moved: The Madrid Health Service (SERMAS) launched a primary-care pilot, integrating access to DxGPT into the clinician's IT platform. The project was framed as a way to shorten the rare-disease “diagnostic odyssey” for patients.
- What early evidence shows: Pre-print evaluations on medRxiv report that the tool can achieve a top-5 accuracy comparable to that of hospital clinicians on complex paediatric case vignettes. However, the studies also highlight that its performance is highly sensitive to the quality of the clinical prompt, and broader LLM studies have flagged inconsistent safety-netting behaviours.
- Governance lessons: The pilot faced scrutiny from Spanish media regarding its contracting and validation protocols. This provides a valuable lesson for the NHS on the absolute need for clear transparency, robust data protection, and a pre-defined evaluation plan from day one.
Implications for the NHS: from enthusiasm to safe, routine use
To bridge the adoption gap, the NHS can take four clear, evidence-based steps.
- Build confidence with use-case pilots: Start where the benefits are tangible and the governance frameworks are already in place. The national guidance for ambient scribing provides a ready-made, safe pathway to introduce AI for administrative tasks. For clinical use, research-grade pilots of differential diagnosis tools in contained settings like genetics or paediatrics can build evidence and trust.
- Train for prompted practice: The low confidence and awareness figures are a direct call for action. The NHS must invest in short, role-specific training that teaches clinicians how to use these tools safely and effectively, covering prompt structure, source verification, and how to document the outputs. The NHS AI Knowledge Repository should be the national home for these patterns and playbooks.
- De-risk via regulatory sandboxes: The UK has a world-class asset in the MHRA’s AI Airlock. This should be the default pathway for testing and de-risking novel clinical AI tools (like differential-generation aids) in supervised conditions, clarifying the evidence requirements before any wider scaling.
- Mandate provenance-first design: The most effective way to address the UK's elevated fear-of-error barrier is to mandate transparency. Procurement frameworks should preference tools that provide clear, auditable citations back to trusted primary sources like NICE guidelines.
A pragmatic adoption framework for ICSs and trusts
| Timeframe | Action |
|---|---|
| Q1 (First 3 Months) | Baseline & Awareness: Audit current (formal and informal) AI use. Run briefings for clinical leaders using the Health Foundation polling data to legitimise interest and frame the challenge. |
| Q2 (Next 3 Months) | Two Focused Pilots: Launch (i) an administrative pilot (e.g., an ambient scribe, following NHSE guidance) and (ii) a clinical pilot in a research setting (e.g., a DxGPT-style tool in a specialist clinic) with pre-registered endpoints. |
| Q3 (Next 3 Months) | Skills & Assurance: Deliver micro-learning on safe prompting and source verification. Enrol any novel medical device tools in the MHRA AI Airlock where suitable. Prepare your DTAC and clinical-safety (DCB0129/0160) artefacts. |
| Q4 (Final 3 Months) | Evaluate & Scale: Publish the results of your pilots (positive or negative) to the NHS AI Knowledge Repository. Only expand the use of tools where accuracy, safety, and productivity metrics have cleared agreed thresholds. |
Success measures should include: The pre/post training confidence delta, awareness of use-cases, time-to-answer, documentation minutes saved, top-5 differential hit-rate (for diagnostic aids), and all safety events.
Risks, limits and how to mitigate them
- Prompt sensitivity & safety-netting: As the DxGPT evaluations show, performance is dependent on input quality. Mitigate by providing training on structured prompting and pairing any AI suggestion with an explicit "seek specialist advice" or "verify with primary source" nudge.
- Over-generalisation from pilots: Mitigate by using the staged evidence requirements of the AI Airlock and NHS guidance. Do not extrapolate the findings from a small, specialist pilot to a general rollout without further validation.
- Governance drift: The Spanish scrutiny of the DxGPT pilot highlights why contracts, data flows, and validation plans must be crystal-clear from day one.
FAQs
- If NHS staff support AI, why aren’t they using it?
- The data shows that UK healthcare professionals report a higher fear of clinical errors (62%), lower confidence, and lower awareness of practical use-cases compared to their peers in France, Germany, and Denmark. These specific barriers must be addressed with training and the adoption of transparent, "provenance-first" tools.
- What’s the evidence on a tool like DxGPT?
- Early academic studies show it can achieve clinician-comparable top-5 accuracy on complex paediatric cases, but its performance is highly dependent on the quality of the input. The public health service in Madrid, Spain, has piloted its use in primary care.
- What national support exists for safe adoption in the UK?
- The UK has a strong set of national guardrails, including the MHRA AI Airlock (a regulatory sandbox), specific NHS England guidance (e.g., for ambient scribing), and the NHS AI Knowledge Repository for sharing lessons and best practices.