The problem is not a lack of information. NICE publishes national guidance, CKS covers primary-care conditions, SIGN provides Scottish recommendations, SmPCs on eMC provide product-specific medicines information, and Trust intranets hold local formularies and pathways. The difficulty is finding and reconciling the right material during a 10-minute consultation or a busy ward round.
The Old Model
A query about a new diagnosis in a patient with renal impairment may require several checks: use NICE or CKS for the management pathway; inspect the relevant SmPC on eMC for the product’s licensed dose and renal thresholds; check the Trust formulary for local restrictions; and use an appropriate interaction checker, SPS advice or pharmacy support where the clinical significance of an interaction is unclear.
Six sources. Six tabs. Six different interfaces. The clinician manually reconciles the outputs and makes a decision — all while the patient waits.
The New Model
Ask a natural-language question. Receive a concise answer with citations to the relevant NICE guideline, CKS topic, SIGN recommendation, SmPC section or local source. Open the cited passage, verify it, and make the clinical decision with the authoritative source in view.
This is not "AI replacing doctors." It is AI replacing the tab-switching, cross-referencing, and manual synthesis that consumes clinical time without adding clinical value. The clinician still makes the decision. The AI compresses the retrieval step.
Why RAG/Source-Grounded AI Is Different from Generic Chatbots
Not all clinical AI is the same. The critical distinction: source-grounded AI (Retrieval-Augmented Generation, or RAG) retrieves information from specified authoritative sources and cites those sources in its response. Generic chatbots (standard ChatGPT, Gemini, Claude) generate responses from training data without guaranteed source specificity or citation.
For clinical use, this distinction is safety-critical. A source-grounded tool answering "what is the NICE-recommended first-line for type 2 diabetes?" retrieves the answer from NICE NG28 and cites it. A generic chatbot may generate a plausible answer based on training data that includes US (ADA), European (EASD), and UK (NICE) guidelines — without specifying which guideline framework it is citing and without guaranteeing it is citing the current version.
NICE itself recognises this evolution. NICE has stated that it is developing guidance for AI methods, evaluating AI technologies for NHS use, and exploring AI tools internally. The NICE Evidence Standards Framework exists specifically to help evaluators and companies judge whether digital health technologies are likely to benefit users and the health system — providing a structured pathway for tools in this emerging category.
A 2026 UCL study published on arXiv demonstrated a RAG system for querying NICE clinical guidelines, achieving 98.7% accuracy with GPT-4.1 while reducing unsafe responses by 67% compared to smaller models — confirming that source-grounded approaches meaningfully improve clinical AI safety.
NICE, CKS, SIGN, SmPCs and Local Pathways: The UK Evidence Stack
Understanding what each source provides is essential for evaluating whether a clinical AI tool has adequate coverage.
NICE Guidelines. Comprehensive national recommendations covering conditions, procedures, and technologies. Evidence-graded. Cost-effectiveness-informed (QALY thresholds). Updated on a rolling basis. The authoritative reference for NHS clinical governance.
NICE CKS (Clinical Knowledge Summaries). Primary-care-focused management summaries. Condition-led — each topic provides a stepwise management pathway from investigation through treatment to referral criteria. More practical and concise than full NICE guidelines. The tool most GPs reach for first.
eMC (SmPCs). Product-led regulatory information for a specific medicine, strength and formulation. Use it for licensed indications, product-specific dosing, contraindications, interactions, adverse reactions, excipients and other authorised particulars.
SIGN (Scottish Intercollegiate Guidelines Network). Scottish evidence-based clinical guidelines. May cover conditions or clinical areas where NICE guidance is absent or where Scottish practice differs. Valuable complementary source, particularly for clinicians in Scotland.
SmPC (Summary of Product Characteristics). Manufacturer's MHRA-approved product information, hosted on the emc. Full adverse event listings, interaction data, dose adjustments by renal/hepatic function, excipient information. The regulatory reference for individual medicinal products.
A clinical AI tool for UK practice should cover relevant NICE, CKS and SIGN guidance, product information from SmPCs on eMC, and clearly identify whether local Trust pathways are included. Source coverage should be explicit rather than inferred from generic claims about UK alignment.
The Emerging UK Tools
iatroX. Free. MHRA-registered, UKCA-marked. Retrieves and synthesises NICE guidelines, CKS summaries, peer-reviewed literature, and SmPC data. But iatroX is not just a search box — it combines clinical AI retrieval with adaptive exam Q-banks, clinical calculators, and CPD documentation. This hybrid model means the clinician interacts with the platform daily across multiple workflows (exam preparation, clinical queries, scoring, CPD) rather than visiting a search tool occasionally.
Praxis Medicine. New entrant. Voi co-founder Douglas Stark. Balderton and Creandum backing. 70 million SEK raised (Breakit, April 2026). Explicitly lists NICE Guidelines, NICE CKS, NHS Digital, and Europe PMC as sources. NHS Website Content API access referenced. Early stage — but credibly positioned.
Medwise AI. UK enterprise deployment. NHS Trust integrations with local policy and formulary content. HRA-listed pilot study comparing AI guideline search against manual intranet search. Available to institutions, not individual clinicians.
Umbil. Publicly positions itself around UK clinical guidance retrieval and workflow outputs such as referral letters, discharge summaries and SBAR handovers. Check its current product documentation for the precise source set and update model rather than assuming coverage from a generic UK-guidance claim.
How to Judge Them
Five evaluation criteria for any clinical AI guideline retrieval tool.
Citations. Does the tool identify the exact source and passage — for example, a NICE recommendation, CKS topic, SIGN section or product-specific SmPC section on eMC? A generic source label is not enough for clinical verification.
Source coverage. Does it cover the relevant NICE, CKS and SIGN guidance? Does it include product information from SmPCs on eMC? Are local Trust pathways included, and are they current for the user’s organisation?
Update frequency. How quickly does the tool ingest guideline updates after NICE publishes a revision? A tool citing an outdated guideline is worse than no tool.
Governance. Is it MHRA-registered? Does it have a clinical safety case (DCB 0129)? What data governance applies to clinical queries?
Usability. Can you use it during a consultation without disrupting the clinical workflow? Is it mobile-optimised for ward use? Does it integrate with anything else you use daily?
Where iatroX Fits
iatroX is built for this retrieval layer — but with clinical education, Q-banks, and calculators attached. Not just a search box, but a daily clinical platform.