NHS AI Adoption Assessment Tool

Adopting an AI tool is rarely a simple yes or no. The same tool can be a sensible choice for one organisation and a serious risk for another, because what matters is the fit between the demands a tool places on its environment and the capabilities of the organisation deploying it. This tool makes that fit explicit. It pairs twelve complexity dimensions (properties of the tool, such as how hard its task is, how its data behaves, and how much it must integrate with existing systems) with twelve readiness dimensions (capabilities of your organisation, such as its governance, data infrastructure, and capacity for clinical oversight). Where a tool's complexity outruns your readiness, you have a gap. The tool identifies those gaps, prioritises them, and returns one of four recommendations: quick win, deploy and monitor, build readiness first, or avoid.

It is a diagnostic instrument, not a prescriptive one. It tells you where to focus; it does not tell you exactly what to do, because that depends on knowledge of your organisation that no standardised tool can hold.

Use this assessment early — while horizon scanning, building a business case, preparing for procurement, or planning a pilot. It is most useful before you have committed, when its findings can still shape the decision rather than justify one already made.

It complements, and does not replace, the formal processes that govern AI in the NHS. It is not a substitute for the clinical safety work required under DCB0129 and DCB0160, a Data Protection Impact Assessment, the Digital Technology Assessment Criteria (DTAC), or regulatory approval of the tool as a medical device. Think of it as the structured conversation to have before and alongside those processes: a single place to surface whether a tool is the right fit for your organisation, and where your effort should go if you decide to proceed.

It is for anyone involved in deciding whether to adopt an AI tool in an NHS setting: clinical safety officers, digital transformation leads, AI programme managers, information governance teams, and procurement. It is most usefully completed by a multidisciplinary team rather than one person, because the questions span clinical, technical, governance, and operational ground that no single role sees in full. Where you do not know an answer, the tool signposts who to speak to — your clinical safety officer, cybersecurity team, information governance lead, or supplier — so that a gap in knowledge becomes a prompt to consult the right person rather than a guess.

This assessment evaluates one tool at a time, but readiness is not independent across tools. The burden on governance, clinical workforce, IT infrastructure, and monitoring accumulates: an organisation that comfortably meets the threshold for a single tool may find that deploying three or four at once exceeds its capacity. Where several tools are in view, run a separate assessment for each and review the readiness scores side by side, paying particular attention to the dimensions where cumulative demand is highest.

Throughout the tool you will see illustrative example answers drawn from eight archetypal tools, spanning the range from administrative to fully autonomous and agentic. They are composites used to show how different kinds of tool tend to score, and to make the reasoning concrete — they are for illustration only and are not assessments of real products.

• FlowCast — Administrative. An operational forecasting tool that predicts demand and capacity — A&E attendances, say — to help plan staffing and beds. It uses no patient-identifiable clinical data and sits outside the clinical decision, anchoring the low-complexity, low-autonomy end of the scale.
• ScribeMate — Administrative in a clinical setting. An ambient documentation assistant that listens to a consultation and drafts the note. It handles patient data and sits inside the clinical workflow, but it records rather than recommends — it does not influence the clinical decision itself.
• RiskSense — Clinical decision support. A risk-prediction and early-warning tool that flags patients at rising risk — of deterioration, a fall, or admission — so staff can intervene. It informs a human decision rather than making one.
• ScanRead — Clinical decision support. A diagnostic tool that interprets medical images and flags suspicious findings for a clinician to confirm. A clear example of decision support: it shapes the decision, but a human makes the call.
• SymptomChat — Clinical decision support. A patient-facing chatbot that gathers symptoms and routes people to the right service or referral. It interacts directly with patients and its language-model behaviour makes it stochastic, but a clinician remains responsible for the clinical decision.
• ClinPilot — Clinical decision support (agentic), agentic. An agentic clinical copilot that not only drafts notes but takes actions in the record — ordering tests, coding — under clinician supervision. It is the key example of agency without autonomy: it plans and acts across multiple steps, but a clinician signs off each time.
• DoseGuide — Bounded autonomous. A bounded-autonomous tool that adjusts a single, well-defined treatment — insulin dose — within a fixed protocol, without a clinician reviewing each adjustment. Autonomous, but on one narrow decision.
• CareAgent — Fully autonomous + agentic, agentic. An autonomous, agentic patient-facing care agent that holds open-ended conversations and acts across many tasks without a clinician reviewing each output. It anchors the highest-scrutiny corner of the framework, where the strictest scoring floors apply.