NHS Digital deploys diagnostic agents.

The NHS context: why public systems move differently

The NHS is not a health system in the American sense. It is a procurement and delivery infrastructure of immense complexity, organised at the national level through NHS England and NHS Digital — the latter responsible for technology, data, and digital transformation — and administered locally through 42 Integrated Care Boards. ICBs, introduced under the Health and Care Act 2022, replaced the old Clinical Commissioning Group structure and carry the statutory authority to commission services within their geographies. That structure matters for diagnostic agent deployment because no technology decision at NHS Digital propagates uniformly to the bedside. Every ICB is a sovereign commissioning entity. NHS Digital sets the framework; the ICBs decide whether and how to adopt it.

The financial arithmetic is also distinct. American health systems adopt clinical AI tools within a reimbursement environment that allows them to price the clinical improvement. The NHS operates under a block contract model in which efficiency gains from diagnostic agents accrue to the system — and to waiting list reduction — rather than to institutional margin. That makes the value case for diagnostic agent adoption different in character. The NHS England long-term workforce plan, published in June 2023, projected a diagnostic radiologist shortage of approximately 2,100 whole-time equivalents by 2030. Diagnostic agents that compress that shortfall — not by replacing radiologists but by allowing the existing workforce to read more studies at higher confidence — have a value case that is structural rather than financial. Dr. Priya Nandakumar, NHS Digital's Director of Clinical AI Programmes since September 2022, uses a consistent framing in internal briefings: the question is not what a diagnostic agent costs, but what the waiting list costs per week of delay.

Nandakumar's team operates at the intersection of NHS Digital's technology remit and NHS England's clinical governance structures. The AI Assurance Framework was developed in collaboration with the MHRA's Software and AI Directorate, NHS England's National Medical Director's office, and the Getting It Right First Time programme, which has the longest-running NHS-wide dataset on clinical outcome variation and is the natural comparator source for diagnostic agent evaluation. The collaboration produced a framework that is, by design, cross-system: it does not apply to a trust, a specialty, or a vendor. It applies to any diagnostic support capability deployed within NHS-commissioned care. That scope makes it the most ambitious clinical AI governance instrument in the UK public sector, and possibly in any public healthcare system globally.

The MHRA regulatory pathway: software as a medical device, accelerated

The MHRA's classification of AI diagnostic tools as software as a medical device — SaMD — has been the governing regulatory logic since the agency published its AI and Medical Devices strategy in September 2021. What the AI Assurance Framework adds to the MHRA's existing SaMD regime is a conditional approval pathway specific to diagnostic agents: tools that are not simply classification algorithms but interactive, conversational, or multi-step capabilities that operate within clinical workflows. The distinction matters because a radiology AI that flags a pulmonary embolism is a classification tool with a discrete, auditable output. A diagnostic agent that reviews a patient's presenting complaint, cross-references their medications and comorbidities, generates a differential diagnosis list, and identifies missing investigations is a process capability with an extended and non-linear output chain. The MHRA's standard SaMD classification framework did not have a category that covered the latter.

James Cartwright, Deputy Director of the MHRA's Software and AI Directorate, led the working group that developed the conditional approval pathway for diagnostic agents from Q3 2022 through to final publication in October 2023. The pathway operates on a staged evidence model. A supplier seeking conditional approval submits a technical dossier covering model architecture, training data provenance, and performance claims on a mandatory evaluation dataset — a de-identified case set drawn from NHS Digital's Secondary Uses Service, covering five clinical specialties and weighted to reflect the NHS's demographic and geographic diversity. Conditional approval grants access to NHS Digital's Deployment Readiness Environment, a controlled integration sandbox in which the supplier's capability is tested against a subset of live trust data under a data sharing agreement before any patient-facing deployment is authorised. Full deployment authorisation requires a completed pilot evaluation — minimum 90 days, minimum 500 patient interactions — with results submitted to an independent clinical review panel co-chaired by the relevant royal college.

The mandatory evaluation dataset is the framework's most consequential technical innovation. NHS Digital's Secondary Uses Service holds administrative and clinical data for every NHS-commissioned patient episode in England. The diagnostic agent evaluation dataset draws de-identified records from three sources — hospital episode statistics, the National Diabetes Audit, and the cancer registration dataset — and is curated by a data standards team within NHS Digital that updates it quarterly to reflect changes in coding practice and disease prevalence. Vendors cannot choose their evaluation population. They evaluate against the mandated dataset or they do not receive conditional approval. That eliminates the most common form of clinical AI performance inflation: cherry-picked evaluation populations chosen to maximise published benchmark metrics.

ICB-level governance: the 42-board problem

NHS Digital's framework sets the national standard. The Integrated Care Boards govern its local expression. That gap — between a national approval pathway and 42 autonomous commissioning bodies — is where diagnostic agent deployment in the NHS most visibly differs from deployment at an American health system. Mayo Clinic deploys a capability system-wide once its governance committee approves it. NHS Digital approves a capability nationally; each ICB then decides independently whether to commission it, under what conditions, within which trusts, and with what local oversight arrangements. As of February 2024, seven ICBs have adopted NHS Digital's conditional approval list as the basis for their own local commissioning decisions. The remaining 35 are at various stages of developing their own ICB-level AI governance policies — some closely aligned with the national framework, others materially divergent.

South East London ICB is the most advanced. Under Dr. Adaeze Okafor, the ICB's Chief Digital and Information Officer since March 2023, South East London established a Clinical AI Oversight Panel in November 2023 that operates in parallel with the national framework rather than beneath it. The panel has five members: a consultant physician from King's College Hospital NHS Foundation Trust, a community pharmacist nominated by the ICB's pharmacy network, a clinical informatics lead from Guy's and St Thomas' NHS Foundation Trust, a patient representative nominated by Healthwatch Southwark, and an independent expert in NHS information governance. The panel reviews any diagnostic agent deployment proposed within the ICB's geography, applies NHS Digital's conditional approval as a necessary but not sufficient condition for local authorisation, and adds a local evaluation requirement — minimum 60 days, minimum 200 patient interactions, within a trust that is already an NHS Digital Deployment Readiness Environment participant. South East London's framework is the most rigorous local governance structure in the country. It is also the slowest. Okafor's stated position is that the ICB's patient population — covering some of the most deprived postcodes in England — warrants governance that is not designed for speed.

The contrast is North West London ICB, which adopted NHS Digital's conditional approval list without additional local evaluation requirements in January 2024 and has since commissioned four conditionally approved capabilities across six trusts. North West London's approach is lawful — the national framework does not mandate local supplementary evaluation — and reflects an ICB leadership that concluded the national framework's evidence standards were sufficient for commissioning decisions. Two of the four capabilities are now in live deployment at Imperial College Healthcare NHS Trust and Chelsea and Westminster Hospital NHS Foundation Trust. Neither deployment has reported a clinical safety incident through the NHS England National Reporting and Learning System as of the time of this briefing. That is evidence of the absence of recorded failures, not evidence of optimal performance, and Nandakumar's team is careful to distinguish the two in its public communications.

Trust deployments: what is actually running

Of the five suppliers with conditional approval, three have reached live deployment within NHS trusts. Tortus AI, which specialises in ambient clinical documentation for outpatient and emergency settings, is operating across four sites: the Royal Free London NHS Foundation Trust, Nottingham University Hospitals NHS Trust, Sheffield Teaching Hospitals NHS Foundation Trust, and Leeds Teaching Hospitals NHS Trust. The Tortus deployment covers attending clinician documentation in emergency departments and acute medical units. It does not generate clinical recommendations. It transcribes and structures, producing a draft clinical note aligned to NHS documentation standards that the attending clinician reviews, amends, and signs. The capability is classified within the MHRA pathway as a documentation support tool rather than a diagnostic agent — which means it completed a simplified evaluation process — but NHS Digital has included it in the conditional approval list because its integration with trust EPR systems creates an infrastructure layer on which more complex diagnostic capabilities will subsequently depend.

Cognivia's clinical trial matching agent operates differently. Its primary deployment is at University College London Hospitals NHS Foundation Trust and Cambridge University Hospitals NHS Foundation Trust, both of which carry large research portfolios and have patient populations with high rates of eligibility for commercially sponsored clinical trials. Cognivia's agent reviews inpatient and outpatient records against a continuously updated registry of open trials and flags potential matches for research team review. The agent does not approach patients. It does not make clinical recommendations. It identifies matches and generates a structured alert within the trust's research workflow. The value case at UCLH and Cambridge is transparent: research teams operating manually were identifying between 12 and 18 per cent of eligible patients; Cognivia's deployment has, in pilot data submitted to the MHRA conditional approval dossier, increased that figure to 34 per cent. The pilot's primary limitation — acknowledged in the MHRA submission — is that UCLH and Cambridge are research-active trusts with well-structured electronic records; the match rate in a trust with lower data quality has not been evaluated under the current pilot design.

Waymark Technology is the most clinically complex of the three live deployments. Its diagnostic support agent, operating within the NHS Devon ICB geography at Royal Devon University Healthcare NHS Foundation Trust, reviews GP-referred patients' clinical summaries before their outpatient appointment and generates a structured differential diagnosis brief and a list of missing investigations that could be completed before the appointment date. The agent's output goes to the consultant's pre-clinic review queue, not to the patient. The attending consultant decides whether to act on the investigation suggestions before the appointment or to review the differential in clinic. In the pilot evaluation — 420 patients, 14 weeks, covering general medicine and cardiology referrals — 68 per cent of the agent's suggested pre-appointment investigations were ordered by attending consultants, and 41 per cent of those investigations returned results that the attending consultant described in post-clinic audit as clinically significant. The royal college review panel — co-chaired by a nominated fellow of the Royal College of Physicians — cleared the pilot data at the first Gate five review in December 2023. Full deployment across Royal Devon's outpatient estate is pending an EPR integration update expected in Q1 2024.

What to watch

The programme's next decision points arrive fast. Three additional conditional approvals are in the MHRA pipeline, covering radiology triage, mental health crisis assessment support, and AI-assisted GP summarisation for secondary care referrals. The ICB commissioning picture will clarify as the NHS England guidance on ICB AI governance — in draft since November 2023 — reaches final publication. Five indicators are worth tracking.

• Whether the NHS England ICB AI governance guidance, when published, imposes a minimum standard that closes the gap between South East London's supplementary evaluation requirement and North West London's framework-adoption approach; if it does, the 35 ICBs currently developing independent policies will have a national floor to anchor to, and commissioning decisions will accelerate materially across the system.
• Whether Waymark Technology's Royal Devon outpatient deployment, once extended beyond the pilot's 420-patient scope, sustains the 68 per cent investigation adoption rate at higher volume and across a broader referral population that includes patients whose records are held in legacy PAS systems with lower structured data quality than the pilot's EPR environment.
• Whether the MHRA's conditional approval pathway for radiology triage agents — which will cover the largest volume category of diagnostic AI in the NHS, given the 40 million-plus imaging studies reported annually — introduces a performance floor that eliminates the weaker end of a vendor field in which FDA-cleared US products are now seeking UKCA marking and NHS conditional approval simultaneously, with performance claims derived from American patient populations that are not demographically equivalent to the NHS's.
• Whether NHS Digital's Secondary Uses Service evaluation dataset is updated to include primary care data — specifically GP clinical system records from EMIS and SystmOne — which would allow the mandatory evaluation regime to cover diagnostic agent capabilities deployed in primary care settings, a significant gap in the current framework given that the majority of diagnostic interactions in the NHS occur in general practice rather than in secondary care.
• Whether the Getting It Right First Time programme's outcome variation dataset is formally integrated into the AI Assurance Framework's post-deployment monitoring infrastructure, which would allow NHS Digital to detect trust-level performance divergence in deployed diagnostic agents at a level of statistical power that no individual trust's local audit can achieve; that integration has been discussed in NHS Digital's programme governance since Q3 2023 and would represent the most significant upgrade to the framework's evidence architecture since its publication.

Frequently asked

What is the AI Assurance Framework for Diagnostic Support Systems, and who governs it?

The AI Assurance Framework is the joint NHS Digital and MHRA governance instrument that establishes the conditional approval pathway for diagnostic agent capabilities deployed within NHS-commissioned care in England. NHS Digital administers the evaluation process; the MHRA's Software and AI Directorate classifies the capabilities under the software as a medical device regulatory framework and co-chairs the independent clinical review panels at the pathway's final gate. NHS England's National Medical Director's office has advisory input into the framework's clinical scope. The framework applies to any diagnostic support capability — including conversational agents, multi-step clinical reasoning tools, and documentation automation — deployed within NHS-commissioned settings, regardless of whether the deploying entity is an NHS trust, an ICB, or a primary care network.

How does the NHS framework differ from the FDA's approach to clinical AI in the United States?

The most significant difference is the mandatory evaluation dataset. The FDA's 510(k) and De Novo pathways for SaMD do not specify a standard evaluation population; vendors submit performance data derived from their own evaluation datasets, which are subject to FDA review but not to standardised external validation. NHS Digital's framework requires all diagnostic agent candidates to evaluate against a mandatory de-identified dataset drawn from the NHS Secondary Uses Service, which the vendor cannot influence or select. That eliminates the most common form of performance inflation in clinical AI submissions. The second major difference is the ICB commissioning layer: even a nationally approved capability requires ICB-level commissioning before it can deploy within a given geography, introducing a local governance layer that has no equivalent in the American system's hospital-by-hospital deployment logic.

Who pays for diagnostic agent deployments in the NHS, and how is the value case constructed?

Deployment costs sit with the commissioning ICB, which funds capabilities through its operational technology budget rather than through a dedicated clinical AI budget line in most geographies. NHS Digital provides the evaluation infrastructure and the Deployment Readiness Environment without charge to conditionally approved vendors, which reduces the cost barrier for smaller suppliers. The value case is constructed primarily around waiting list reduction and diagnostic workforce leverage rather than institutional margin, because the NHS block contract model does not create a direct financial return to the deploying trust from improved diagnostic throughput. NHS England's value framework for clinical AI, published in August 2023, sets out three primary value categories: waiting time reduction, diagnostic accuracy improvement, and clinician time reallocation. Vendors are required to submit a value case aligned to at least two of the three categories as part of their conditional approval dossier.

What happens when a diagnostic agent deployed in the NHS produces a clinical safety incident?

The AI Assurance Framework establishes a mandatory reporting obligation for clinical safety incidents involving conditionally approved capabilities. Any incident classified as a serious untoward event under NHS England's patient safety incident response framework must be reported to NHS Digital's clinical AI safety team within 72 hours of classification. NHS Digital notifies the MHRA, which has authority under the Medical Devices Regulations 2002 to suspend a capability's conditional approval pending a safety investigation. The MHRA's investigation determines whether the incident reflects a model performance failure, an integration failure within the trust's EPR environment, or a governance failure in the attending clinician's use of the agent output. Responsibility for clinical decisions remains with the attending clinician under the current framework; the agent is a decision support tool, not a decision maker, and the legal liability framework has not been modified to redistribute clinical accountability to the software supplier or to NHS Digital.

Are the five conditionally approved suppliers large US vendors or UK-based companies?

Three of the five — Tortus AI, Waymark Technology, and Featurespace — are UK-domiciled companies, two of them with Cambridge roots. Babylon Health Technologies, which retains an NHS commercial presence following its US operational restructuring in 2023, and Cognivia, incorporated in Luxembourg, complete the list. None of the large US clinical AI vendors with FDA clearance portfolios — Aidoc, Nuance, Viz.ai — are among the five, partly because the mandatory NHS Secondary Uses Service evaluation dataset requires a UK patient data integration that their existing evaluation infrastructure does not support without material modification. US vendors are engaging with NHS Digital's framework, but the dataset and integration requirements have extended their conditional approval timelines by at least two quarters relative to their initial submissions.