Executive summary
The landscape of AI clinical decision support (CDS) in the UK is evolving at a rapid pace. For clinicians, the goal is no longer just to find information, but to get a trustworthy, provenance-first answer, at the point of care, without adding to their cognitive load. This has led to a clear split in the 2025 toolkit.
The foundational reference layer remains the authoritative, nationally-recognised sources: NICE CKS for primary care pathways, the BNF for prescribing, and institutionally-licensed compendia like BMJ Best Practice and DynaMedex. Alongside this, a new "AI front-door" has emerged, featuring conversational Q&A tools like the UK-centric iatroX, the US-focused OpenEvidence, the UK-based Medwise AI, and differential diagnosis builders like Glass AI.
For NHS organisations, the path to safe adoption is now clearly defined. Any new tool must be evaluated against the DTAC procurement baseline, the DCB0129/0160 clinical safety standards, and the evidence expectations of NICE, with novel AI devices being tested in the MHRA AI Airlock. This guide provides a practical map of the landscape and a playbook for how to adopt these powerful new tools safely.
The UK CDS landscape at a glance
To navigate the market, it's helpful to categorise the tools by their primary job:
- Reference backbones (the "ground truth"): These are the core, editorially-maintained sources you rely on.
- NICE CKS: The first stop for concise, primary-care pathways.
- BNF/BNFC: The non-negotiable standard for UK prescribing.
- BMJ Best Practice: Excellent for structured condition overviews and its unique comorbidity manager, freely available to NHS staff via OpenAthens.
- DynaMedex: A leading evidence-based summary tool, often bundled with the Micromedex drug database.
- AI assistants (the "fast front door"): These tools use AI to search, synthesise, and provide conversational answers with citations.
- iatroX: A free, UK-centric platform with a citation-first Q&A tool (Ask iatroX) and a Knowledge Centre built on a gated library of peer-reviewed research and UK-accepted guidance (like SIGN), all designed to complement the core NHS sources.
- OpenEvidence: A free, AI-powered reference tool for verified clinicians, primarily focused on the US and global peer-reviewed literature.
- Medwise AI: A UK-focused clinical search tool that includes the ability to retrieve local NHS Trust guidance.
- Glass AI: A specialised AI tool for scaffolding a differential diagnosis and drafting a management plan.
What to use when: Use the BNF for a quick prescribing check. Use CKS or BMJ Best Practice for a concise pathway. Use a tool like iatroX or OpenEvidence for a rapid, cited synthesis of the evidence. Use iatroX Brainstorm or Glass AI to help structure a complex differential diagnosis.
Taxonomy: from search to “show me in workflow”
The evolution of CDS can be seen as a five-step journey:
- Find: Classic keyword search in trusted sources like CKS, BNF, BMJ Best Practice, or DynaMedex.
- Explain: AI-grounded Q&A that provides a natural language answer with visible citations (e.g., iatroX, OpenEvidence, Medwise).
- Decide: Structured AI suggestions for a differential diagnosis, red-flags, or investigations (e.g., iatroX Brainstorm, Glass AI).
- Act: Embedding these tools directly into the EHR via SMART on FHIR or triggering CDS Hooks at the point of ordering or diagnosis.
- Learn: Automatically capturing these learning moments for a CPD portfolio (a core feature of iatroX).
Safety & assurance (UK)
Before any UK CDS tool is deployed, it must pass a clear set of governance hurdles:
- DTAC (Digital Technology Assessment Criteria): The mandatory baseline for procurement, covering information governance, security, clinical safety, usability, and interoperability.
- Clinical safety cases (DCB0129/0160): The manufacturer must provide a DCB0129 safety case, and the deploying NHS organisation must complete a DCB0160 safety case.
- NICE evidence standards: For tools making clinical claims, their evidence must align with the NICE Evidence Standards Framework (ESF) or the Early Value Assessment (EVA) pathway.
- MHRA regulation: Novel AI tools that provide a diagnosis or drive treatment may be classified as a medical device (AIaMD) and are subject to MHRA regulation, including pathways like the AI Airlock sandbox.
How the major tools compare
| Task | Best-fit Tools | Strengths | Watch-outs |
|---|---|---|---|
| Rapid Guideline Lookup | NICE CKS, BMJ Best Practice, DynaMedex | Authoritative, structured pathways | May require a subscription or OpenAthens login |
| Prescribing & Interactions | BNF/BNFC, SPS | The definitive UK national standard | AI tools must be verified against these |
| Cited Q&A / Synthesis | iatroX, OpenEvidence, Medwise | Provenance-first, conversational, fast | Must ensure citations are UK-applicable |
| Differential & Red-Flags | iatroX Brainstorm, Glass AI | Mitigates clinician's anchoring bias | Keep clinician-in-the-loop; verify against primaries |
| Learning & CPD | iatroX (CPD/Quiz), BMJ Learning | Captures reflections at the point of care | Must maintain documentation for appraisal |
Technical pattern for accuracy
The safest AI clinical decision support tools use this common architecture:
- Gated corpus: A "walled garden" of trusted, version-controlled sources. For iatroX, this is a library of UK-accepted guidance and peer-reviewed research.
- Hybrid retrieval: A sophisticated search that combines keywords (lexical) with meaning (embeddings) to find the most relevant passages.
- RAG generation: The AI is forced to generate its answer only from those retrieved passages and must provide the citations.
- Observability & abstention: The system logs all its actions for audit and, crucially, is programmed to "abstain" or say "I don't know" if the evidence is weak or absent.
High-value use-cases to pilot
- Primary care antibiotic choices in pregnancy: KPIs: Time-to-answer, adherence to national guidance, safety events.
- Abnormal LFTs in adults: KPIs: Appropriateness of first-line investigations, unnecessary repeat tests avoided.
- Chest pain triage (non-ACS): KPIs: Correct risk tiering, unnecessary referrals avoided.
- CPD capture: KPIs: Number of CPD entries per week, quality of appraisal-ready reflections.
Integration & workflow
The future of CDS is not in a separate app, but inside the EHR.
- Single Sign-On: Use NHS login or OIDC to reduce login friction.
- SMART on FHIR: A secure standard that allows tools like iatroX or BMJ Best Practice to be launched from within the patient record.
- CDS Hooks: A "trigger" standard that can push a relevant card (e.g., a dosing reminder from the BNF) to the clinician at the exact moment they are prescribing.
Risks & Mitigations
- Automation bias: Mitigate by mandating a "human-in-the-loop" workflow and using tools that show clear uncertainty.
- Hallucination / outdated guidance: Mitigate by only using RAG-based tools with citations to gated, trusted sources. Always check the date stamp.
- Fragmentation: Mitigate by choosing a "front-door" tool (like iatroX) that helps you navigate and link to your core references, rather than adding another silo.
- Data protection: Mitigate by enforcing data minimisation. Never paste patient-identifiable information into free-text prompts on non-approved tools.
FAQs
- Is AI CDS a medical device?
- It depends on its intended use. If it simply retrieves information (like a book), it may not be. If it provides an autonomous recommendation or diagnosis for a specific patient, it is likely a Software as a Medical Device (SaMD) and must follow MHRA rules.
- Do I still need my references if I use AI?
- Yes, absolutely. You should use AI as a fast front-door to find the right information in NICE CKS, the BNF, BMJ Best Practice, or DynaMedex. You must always document the primary-source citations.
- How do we start a pilot without overwhelming clinicians?
- Run a 90-day pilot focused on a single, clear metric (e.g., time-to-answer for common queries). Publish the results internally before scaling.