NHS AI is ready. Our infrastructure isn’t. Here’s what has to change.
Every radiology conference in the last three years has promised the same thing: AI is coming, and it will transform diagnostic imaging. The tools exist. The evidence is growing — and so is the case for AI-driven workflow in radiology. NHS Grampian reduced the time to lung cancer treatment by nine days using chest x-ray AI. Early-stage detection increased by 27%. These are not hypothetical outcomes — they are documented, peer-reviewed results from NHS sites that managed to get AI into clinical deployment.
The question that nobody at those conferences answers directly is this: why haven’t you?
Because for most trusts, the honest answer is that the infrastructure to support AI adoption at scale simply isn’t in place. And until it is, every promising AI tool a trust evaluates will follow the same arc — a compelling pilot, a procurement decision, and then months of stalled deployment while the foundational work catches up.
The Two Walls Every NHS AI Project Hits
Independent analysis of NHS AI adoption, including the 2025 UCL Partners and Health Foundation report, consistently identifies the same two systemic barriers: data preparation and integration complexity.
The data preparation problem is more acute than most people outside radiology research realise. To validate an AI tool for clinical use, a Trust needs a compliant anonymised dataset — often thousands of studies, pixel-level anonymised to UK GDPR standard, with a full audit trail for ethics review. Manual anonymisation of a 6,000 x-ray validation dataset, at ten minutes per study, takes over 1,000 hours.
That is 25 working weeks of clinical or research staff time, consumed by a process that produces no diagnostic value whatsoever. BriX — Rosenfield Health’s bulk DICOM anonymiser — processes the same 6,000 studies in approximately four hours. Ethics review cannot begin until the dataset is ready. AI deployment cannot begin until ethics review is complete. A promising AI tool sits unused for six months not because it failed its evaluation, but because the data pipeline that should have fed that evaluation was never built.
The integration problem compounds this. Each AI tool requires its own connection to a trust’s PACS system. That connection costs approximately £5,000 in technical resource and vendor coordination. A trust deploying five AI tools — a modest ambition for any serious radiology AI programme — faces over £25,000 in integration costs before a single radiologist has seen a result. Multiply this across the vendor relationships, the IT coordination cycles, and the ongoing maintenance burden, and it becomes clear why so many trusts remain in perpetual pilot phase. The economics of building a proper AI program, tool by tool, are punishing.
What Scalable NHS AI Infrastructure Actually Looks Like
The NHS 10-Year Plan commits £600 million to anonymised health data infrastructure for AI. That is a significant signal. But the capital investment only creates value if Trusts have the operational infrastructure to use it — the ability to prepare compliant datasets quickly, and the ability to connect AI tools to clinical workflows without rebuilding the integration from scratch every time.
Scalable AI infrastructure has three characteristics. First, it separates data preparation from clinical workflows, so anonymisation happens in bulk at processing speed rather than manually by clinical staff. Second, it creates a single integration layer between PACS and AI tools — a vendor-neutral orchestration platform such as PAIP — so each new tool connects through the same point rather than requiring its own bespoke connection. Third, it is vendor-neutral, so trusts retain the freedom to adopt best-of-breed AI tools from any vendor rather than being locked into the AI portfolio of their PACS supplier.
None of these characteristics require replacing existing PACS infrastructure. That is a misconception that slows procurement decisions considerably. A properly designed AI adoption platform sits alongside existing PACS and RIS systems, not instead of them. The question for PACS Managers and Clinical AI Leads is not “do we have to rebuild our infrastructure?” but “do we have a layer that connects what we have to the AI tools we want to deploy?”
The Cost of Waiting
NHS England’s AI guidance is explicit that DICOM metadata — even after basic identifier removal — can enable patient re-identification. Pixel-level anonymisation is a compliance requirement, not a preference. Trusts producing AI training datasets through manual processes or basic anonymisation scripts are carrying legal exposure that will eventually be audited.
More pressingly, 1.7 million patients are currently waiting for diagnostic tests across the NHS. AI tools that could support worklist prioritisation, triage, and early detection of diagnostic errors in radiology exist and are validated. The bottleneck is not clinical evidence. It is infrastructure.
The trusts that will lead NHS AI adoption in the next three years will not be the ones that evaluated the most AI tools. They will be the ones that built the infrastructure to move from evaluation to clinical deployment in days rather than months — and that treated data preparation and AI orchestration as foundational investments rather than afterthoughts.
The tools to do this exist now. The question is whether procurement decisions treat AI infrastructure with the same seriousness as the AI tools that depend on it.
The content on this page is provided by the individuals concerned and does not represent the views or opinions of RAD Magazine.
