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

Every clinical rule took a specialist clinician weeks to write.

Manual creation of business rules took four to six weeks per clinical area. It required specialist clinicians to read and interpret hundreds of pages of guidance, and it was inconsistent and bottlenecked by clinician availability. Those clinicians billed a minimum of £500 a day, which made scaling the product across new disease and treatment areas prohibitively expensive.

Automating the process with LLMs introduced its own constraints. Outputs had to be clinically guideline-aligned, auditable, and demonstrably safe for a regulated environment, with strong provenance, reproducibility, and clinician sign-off built in rather than bolted on.

Three constraints were non-negotiable:

• 4-6 weeks of manual work per clinical area
• Specialist clinicians billed at £500+ a day
• Outputs had to be auditable and guideline-aligned

A RAG pipeline that turns guidelines into audited rules.

The Fractional Head of AI designed a retrieval-augmented generation pipeline built around Mistral 7B to extract, summarise, and transform clinical guidelines into structured business rules. Retrieval, grounding, layered validation checks, and human-in-the-loop review were integrated throughout, so clinicians reviewed and signed off on generated outputs instead of authoring rules from scratch.

Governance was built into every step: source citation, evidence linking, rule-by-rule traceability, and full audit trails, plus validation steps equivalent to unit tests and formal clinical sign-off protocols. A modular orchestration framework let the pipeline absorb new models and capabilities quickly, and the RAG architecture supported ongoing ingestion of new guidelines while referencing existing rules to avoid duplication and drift.

Performance monitoring, quality metrics, and feedback loops refined both intermediate outputs and final rule quality without adding to clinician workload. Tailored upskilling sessions gave clinicians, product managers, and engineers hands-on training in prompt engineering, RAG principles, validation methods, and the safe clinical use of LLM outputs.

Weeks of clinician time collapsed into a same-day review.

The pipeline cut rule creation time by 97% and saved more than £2 million across 170 clinical areas, while keeping every rule traceable, evidence-grounded, and clinically signed off. It gave the business a reusable engine for scaling clinical decision support into new disease and treatment areas, and a workforce that trusts and understands the AI doing the work.