Wales already owns one of the best health-data assets in Europe and has no strategy worthy of it. This chapter supplies one: data as a national asset with citizen-governed access, and AI adopted the only way it safely scales — through a published assurance framework, priority use cases chosen for clinical time returned, and the discipline to buy and adapt rather than build.
Start from strength: SAIL
The SAIL Databank at Swansea University has linked, anonymised Welsh health data and a two-decade international research track record. The blueprint’s population health data platform (layer 4) is built with SAIL, not beside it: SAIL-class governance for research access, the national platform for operational analytics — waiting lists, capacity, outcomes — feeding the public dashboard.
Three data rules, applied nationally:
- Data lives where care happens; analysis comes to the data. The federated care record is a query, not a warehouse. Population-scale analysis runs in trusted research environments with published access decisions.
- Secondary use is citizen-governed. A national transparency register of every data use; opt-out honoured end to end; the patient council (Citizens) sits on the access panel. Care.data’s lesson is permanent: consent and communication first, or the asset is forfeit for a decade.
- Open data by default for system data. Waiting times, performance, incidents, spend — published as machine-readable datasets under open licence, the way this site publishes its own.
AI: assurance, then adoption
NHS Wales will adopt AI. The only question is whether it does so deliberately, or vendor by vendor through the side door. The blueprint’s answer is a national clinical AI assurance framework, run by the standards body’s clinical informatics function:
- A public register of AI in clinical use — every deployed model, its intended use, its evaluation evidence, its named clinical safety owner.
- Assurance gates aligned to existing law and standards — UK medical-device regulation, DCB0129/0160 hazard assessment, and post-market surveillance with published incident reporting.
- Local validation before local use. A model assured nationally is still evaluated on the deploying board’s population before go-live. Bias found is bias published.
Priority use cases are chosen by one criterion: clinical time returned per pound. On today’s evidence that means ambient documentation (letting clinicians talk to patients instead of typing), imaging triage in the diagnostics backlog, and waiting-list validation — not chatbots, and not a bespoke Welsh foundation model. Wales buys and adapts; it contributes evaluation evidence upstream; it builds only what no market will sell it.
The governance
One accountable owner: the standards body’s Chief Clinical Informatics Officer holds the AI register and the assurance framework; the citizen access panel holds secondary use; Audit Wales audits both.
Wales has been here before. The author served as Technical Lead of the Welsh Government’s Commission for AI in Health and Social Care and wrote the national AI adoption strategy for the Welsh NHS. It was never implemented. This chapter is its successor — placed where implementation can actually happen: inside the operating model, attached to the funding gates, owned by named roles.