Add us as a preferred source
Vicky Rothwell, consultant enterprise architect at Aire Logic, says that rushing to adopt AI in healthcare can risk amplifying the very problems it’s meant to solve.
Every national health system in the world is talking about artificial intelligence. The potential is genuinely transformative as it could mean earlier diagnosis, reduced administrative burden, smarter resource allocation and more personalised care. Yet for every headline about an AI model outperforming a radiologist, there is a quieter, less glamorous reality playing out inside health organisations. One that is defined by incompatible systems, inconsistent data and a patchwork of local workarounds that have accumulated over decades.
This is the gap that rarely gets discussed at conferences or in ministerial announcements. AI models are only as good as the data they consume. Feed them fragmented, duplicated, or poorly governed information, and the outputs will be unreliable at best and dangerous at worst. In a clinical setting, that is not merely an inconvenience – it is a patient safety risk. Before any health system can meaningfully adopt AI, it needs to confront a more fundamental question: is our digital architecture actually fit for purpose?
What we found in Wales
When the Welsh government and Digital Health and Care Wales (DHCW) commissioned the design of a national target architecture, the brief was ambitious. It wanted to unify the country’s digital health estate and create a foundation capable of supporting future technologies, including AI. The engagement spanned every NHS Wales organisation, and what emerged was a picture that will be familiar to anyone working in health IT across the UK and beyond.
Patient information was scattered across hundreds of applications – more than 850 in total, with more than 300 holding patient data, and many of them bespoke, undocumented or disconnected. There was no single source of truth. Coding schemes and data standards varied between organisations, interoperability and HL7 FHIR implementation were inconsistent, and data lineage was not always traceable. Although the main care record has been national across NHS Wales for over a decade, important elements do not flow between care settings, forcing clinicians into manual workarounds and re-keying of information. Crucially, there was no national framework for AI governance, which meant no shared guardrails for ethics, testing or deployment.
None of this is unique to Wales. Fragmentation is the defining characteristic of health IT estates everywhere. The difference is that Wales chose to address it systematically rather than bolt AI onto a broken foundation.
The approach taken in Wales offers a model that other systems would do well to study. Rather than starting with AI use cases and working backwards, the programme embedded AI readiness into the architecture itself. That meant tackling the prerequisites such as data quality, interoperability, standards, security and governance.
At the heart of the design is the Welsh Health Data Space – a unified, domain-specific data layer aligned to FHIR and European Health Data Space standards, with built-in pseudonymisation, metadata services and quality controls. Alongside it, the Welsh Health Information Space replaces hundreds of bespoke point-to-point interfaces with reusable national services and standardised APIs. An open architecture standards framework ensures every system communicates in a consistent way, with shared data dictionaries, terminology mappings, and a national API catalogue.
For AI specifically, the architecture includes secure, nationally governed research environments where pseudonymised datasets can be explored in isolated compute environments with full audit trails. This allows clinicians, researchers and analysts to develop and validate AI models responsibly – without exposing identifiable data.
Governance is not a barrier – it’s an enabler
One of the most important design decisions was to build a structured AI readiness model directly into the architecture, delivered across two phases. The first establishes a national AI delivery framework covering ethical guardrails, assurance standards, evaluation criteria, and clear processes for testing and approval. The second enables deployment at scale, with standardised rollout patterns, continuous monitoring, operational oversight, and consistent risk controls.
Too often, governance is treated as something that slows innovation down. The Wales programme demonstrates the opposite: by creating a transparent, repeatable national process for evaluating and deploying AI, organisations can move faster and with greater confidence. It removes the need for each Health Board to reinvent the wheel, and it gives clinicians and citizens assurance that AI tools have been properly tested before they touch patient care.
Perhaps the most underrated achievement of the programme is a cultural one. Wales established its first-ever architecture community of practice, bringing together 20 to 30 architects from every Health Board and Trust to co-design the future state. This wasn’t a top-down imposition – it was genuine, sustained collaboration that ensured the architecture reflected real operational needs.
This matters enormously for AI adoption. Successful deployment of AI in healthcare isn’t just a technical challenge; it requires buy-in from frontline staff, shared standards across organisational boundaries, and a collective commitment to data quality. You cannot achieve that through procurement alone. You achieve it by building communities, sharing knowledge, and designing together. The Wales experience suggests that investing in collaborative governance structures may ultimately accelerate AI adoption far more than any technology investment.
What this means for the rest of the sector
Wales is now positioned as one of the first UK nations to have embedded AI readiness into its national health architecture from the ground up, rather than retrofitting it. The ten-year roadmap is realistic about the scale of transformation required, but the foundations are in place: a unified data blueprint, consistent standards, secure research environments, and a structured governance model.
For health systems elsewhere, the message is clear. AI in healthcare is not primarily a technology problem, it is an architecture, data and governance problem. The organisations that will get the most from AI are not those that rush to pilot the latest model. They are the ones doing the unglamorous, essential work of unifying their data, standardising their interfaces, and building the governance structures that make safe adoption possible. That work takes time, political will, and sustained investment – none of which make for exciting headlines. But it is the only path to AI that is safe, equitable, and scalable.
The future of AI-powered healthcare is enormously exciting. But it starts with foundations, not algorithms, and the sooner the sector accepts that, the sooner patients will benefit.
