The NHS already has AI tools that work in the narrow technical sense: validated on benchmark datasets, CE or UKCA marked, with published sensitivity and specificity data. What it has far fewer of are AI tools absorbed into routine clinical practice in a way that demonstrably improves outcomes or frees clinical capacity at scale. Closing that gap is where most of the difficult problems in NHS AI lie, a set of problems that technical performance data alone cannot address.
This post is a companion to AI in the NHS: from potential to practice, which covers the regulatory landscape and systemic barriers. Here we look at what it takes for an AI diagnostic tool to change a clinical pathway, and what the evidence from NHS deployments tells us about where the conditions for that are being met.
When evaluating an AI diagnostic tool, the natural question is whether it performs as well as or better than the current standard. In practice, that comparison is often set against a theoretical baseline, rather than reflecting what NHS patients are receiving.
DVT diagnosis is a useful illustration. NICE guidelines require diagnosis within four hours of presentation, but Freedom of Information data from NHS Trusts puts the average waiting time at 3.3 days. A tool evaluated against “consultant radiologist performing immediate ultrasound” will be measured against a standard that a substantial proportion of patients do not receive. Evidence measured against the real NHS baseline, rather than an idealised standard of care, is harder to generate but considerably more persuasive to commissioners.
The clearest current example of AI changing an NHS diagnostic pathway at scale is radiology. The NHS performs more than seven million chest X-rays a year against a 29% shortfall in radiologists. AI tools that triage and prioritise scans by urgency, flagging high-risk cases for immediate review, address a measurable system pressure.
NHS England’s AI Diagnostics Fund, a £21 million programme, has deployed chest X-ray AI tools across more than 40 NHS trusts and six imaging networks. At NHS Grampian, the time between chest X-ray assessment and commencement of lung cancer treatment fell by nine days on average. A further £30 million announced in July 2026 is intended to extend AI chest X-ray tools to every NHS trust in England by 2029.
Several features of this deployment explain why it has worked where others have stalled:
Most AI diagnostic tools fall into the early-use category of NICE’s HealthTech Programme: technically viable, with some clinical evidence, but lacking the NHS-based comparative data that routine-use guidance requires. The gaps NICE consistently identifies in submissions are avoidable:
NICE publishes evidence generation plans alongside early use assessments. For companies approaching evaluation, reading those plans for related technologies in the same clinical area gives a clear signal of what will be expected.
A Class II medical device that has received UKCA or CE marking has been through a conformity assessment process involving thousands of pages of clinical testing, safety data and documentation. NHS trusts nonetheless frequently ask companies for additional local pilots before adoption, applying a scrutiny standard they would not apply to a licensed medicine entering the prescribing pathway.
The gap has two roots. Clinical and managerial staff are more familiar with the pharmaceutical regulatory pathway than the medical device pathway. And unlike medicines, AI diagnostic tools require active procurement decisions, workflow integration, staff training and IT system changes: a trust that understands a technology’s regulatory standing may still lack the capacity or budget to implement it. The MHRA’s updated regulatory framework for medical devices, being implemented through 2026, aims to create clearer, more proportionate routes that NHS procurement teams can interpret. How much it reduces the implementation burden in practice will become apparent as trusts encounter it.
AI tools do not stay static after deployment. Models can be updated, performance characteristics can shift across patient populations, and clinical context changes. Governance frameworks designed for fixed medical devices- validate once, approve, deploy- are not adequate for systems that can change after they are in use. The NHS does not yet have a settled approach to this.
A growing gap separates NHS organisations building AI infrastructure and digital maturity from those that are not. AI benefits risk accruing to organisations already best positioned to implement them, while trusts serving areas of higher deprivation fall further behind. Central funding programmes such as the AI Diagnostics Fund address this partly, but only where deployment is actively targeted at organisations with the lowest baseline capability.
AI diagnostic tools have been documented to perform less well for non-white patient populations, reflecting skewed training datasets. In a system where the 10-year plan explicitly commits to reducing health inequalities, deploying tools without equity validation creates both a clinical risk and a regulatory one: NICE’s evaluation framework increasingly requires diversity analysis of study populations, and the absence of it is becoming a material barrier to recommendation.
For companies developing clinical AI, the evidence package that reaches NHS scale needs to answer three questions beyond technical accuracy:
Companies that address these questions during evidence programme design, rather than after a conditional recommendation is received, are better placed for both NICE evaluation and the procurement conversations that follow.
The chest X-ray programme and the NHAP’s inclusion of AI tools for cancer histopathology show what is achievable when centralised funding, a defined clinical problem and a measurable pathway outcome align. Those examples are instructive precisely because the conditions that made them work were designed in from the start, not retrofitted after the technology was built.
The 10-year plan sets an ambition for AI to be seamlessly integrated into most clinical pathways by 2035. The distance between that ambition and current adoption is not primarily a technology gap. For most AI diagnostic tools, the barriers are evidentiary, operational and structural. That is where industry and the NHS both have work to do.
Explore our latest thinking, event updates and industry insights to stay informed.
Record NHS mental health spending, a new cross-government strategy and the neighbourhood health shift are reshaping commercial landscape for pharma and medtech
NICE has launched a new healthtech evaluation pathway and a mandatory funding programme. For most medtech companies, the harder work still lies beyond approval.
Category H, the SNF and ICB prescribing pressures are landing at once. Here's what it means for dispensing practices and pharma partners.
.svg)
