Clinical rule creation cut by 97%.
£2m+ saved.

Every rule took a clinician a month to write, at £500 a day.

Clinical decision support tools work from rules extracted out of dense guideline documents such as NICE and SIGN. Writing those rules was a manual job: clinicians read the guideline text, interpreted the nuance, and turned it into machine-readable logic. It took 20 to 30 clinician days per treatment area, at £500 a clinician day — £10,000 to £15,000 for a single rule set.

That cost made scaling the CDS suite across therapeutic areas slow and prohibitively expensive. It also introduced inconsistency: different clinicians interpreting the same dense guideline text produced different rules. Any automated alternative had to stay auditable, evidence-grounded and acceptable to clinical governance — and technical, product and clinical teams had no prior experience with AI-assisted rule generation to draw on.

Three constraints were non-negotiable:

• 20-30 clinician days and up to £15,000 per treatment area
• Subjective interpretation produced inconsistent rule sets
• Had to stay auditable and guideline-aligned for regulated clinical use

A RAG pipeline that turns guidelines into auditable rules.

The team built a retrieval-augmented generation pipeline on Mistral-7B. NICE and SIGN guideline documents were ingested, chunked, embedded and indexed in a FAISS vector store. Retrieval pulled the relevant sections of a guideline, which were fed to the model with structured prompts to generate rule logic — with the source passages extracted alongside every generated rule so clinicians could check provenance in seconds.

A validation layer checked units, numerical ranges and internal contradictions before anything reached a clinician. Clinicians’ role shifted from writing rules to validating them: reviewing each generated rule against its highlighted source text, then approving, editing or rejecting it. Safety and governance were built in throughout — source citation on every rule, a full retrieval audit trail, grounding checks against hallucinated logic, rule-by-rule traceability, and mandatory clinical sign-off with evidence links.

The pipeline was built as a modular orchestration framework, so ingestion, indexing, retrieval, generation, validation and review could each be swapped out independently — new LLMs or retrieval methods slotted in without a rebuild, and updated guidelines could be re-ingested with affected rules automatically reprocessed. Monitoring covered retrieval quality, rule consistency and clinician override rates, feeding back into retrieval parameters and prompts. Prompt engineering, RAG architecture and AI product management training was delivered across clinical, engineering and product teams to build confidence in the new workflow.

97% faster, and £2 million saved.

Evidence-grounded retrieval and structured generation improved the consistency and accuracy of the rules themselves, while the audit trail and mandatory sign-off kept clinical and regulatory confidence intact. The reusable pipeline let the CDS suite scale across multiple disease and therapeutic areas in days rather than months, and the upskilling built AI literacy across clinical, engineering and product teams that outlasted the project itself.