A UK firm trained a “sovereign” NHS triage AI that it says rivals Claude, at a fraction of the cost

The NHS does not lack for AI pitches. What it has lacked is one that keeps patient data inside the country. A British software firm says it now has it.

OneAdvanced, the Birmingham-based SaaS company whose software touches more than 40 million NHS patients a year, has launched what it bills as the UK’s first private, sovereign healthcare large language model trained on NHS primary-care data.

The model, called Care Navigator, was built with Nvidia and trained on pseudonymised, real-world patient triage requests submitted through OneAdvanced’s Patchs online-consultation platform, which handles around 500,000 patient interactions a month.

Its job is triage: detecting the clinical topic in a patient’s request so the system asks the right follow-up questions and routes people to the right care faster, which OneAdvanced says can cut the resource waste that clogs general practice.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!

The selling point is where it all happens. The model’s weights, fine-tuning, hosting, and inference sit entirely within the UK, with data stored and governed under UK law, an answer both to NHS rules on data residency and to wider unease about sending health records to US-controlled clouds.

It is also small. Built on Nvidia’s open Nemotron models, the 9-billion-parameter Care Navigator is a fraction of the size of frontier systems, and OneAdvanced says it ran at up to 150 times lower inference cost while outperforming leading models, including Anthropic’s Claude, in a systematic evaluation.

Its boldest claim is that on a benchmark categorisation task the model scored significantly higher than a GP control group, though that is a narrow test of clinical-topic detection, not a contest of medical judgement. Unlike a general chatbot, it is meant to improve in use: corrections made by GPs become fresh training data.

“The future of AI in healthcare will not be built on generic models,” said Simon Walsh, chief executive of OneAdvanced. “In healthcare, accuracy, governance and context matter more than anything.”

Nvidia, which supplied the underlying models and infrastructure, framed the work as a template for regulated industries.

“The UK has a significant opportunity to lead in the development of sovereign AI for highly regulated and mission-critical sectors such as healthcare,” said Anthony Hills, the chipmaker’s director for UK and Ireland, adding that the pilot showed “the difference organisations with deep NHS expertise and context can make.”

The launch lands in the middle of Britain’s sovereign-AI moment. It follows a London Tech Week studded with sovereignty pledges, the unveiling of Cosine’s home-grown frontier model, and a wave of startups, from Deliverance AI to Nebius, pitching governed, in-country AI to regulated buyers.

Healthcare is the most sensitive test of the idea: the data is intimate, the regulation heavy, and the appetite for cost savings enormous. OneAdvanced is not a newcomer to that world, with 35 years behind it, software in more than 4,000 GP practices and over 160 NHS trusts, and a hand in 85 per cent of NHS 111 services.

There are caveats.

This is a company announcement built on a pilot, the “first” claim is OneAdvanced’s own, and the accuracy and cost figures, including the result against GPs, are self-reported rather than independently audited. The company’s own materials are inconsistent on exactly which Claude models it tested against, citing Sonnet and Opus in one place and Sonnet and Haiku in another.

Training AI on patient data, even pseudonymised, is also the territory where NHS data projects have drawn scrutiny before, and a model that learns continuously from clinicians will have to prove it stays safe as it does.

Still, the pitch is shrewdly matched to the moment: a smaller, cheaper, UK-hosted model aimed at a health service desperate to do more with less, without handing its data to anyone abroad.

Whether Care Navigator graduates from pilot to national triage tool is the real test. If it does, it becomes a data point for the argument that sovereign AI need not mean second-best.