OneAdvanced has completed a pilot with NVIDIA to develop and validate a sovereign AI model for NHS care navigation. The model was trained on pseudonymised online consultation requests from NHS primary care patients.
The system, called Care Navigator LLM, is designed to support AI-assisted patient triage and care navigation in primary care. Pilot testing found it achieved significantly higher accuracy than a GP control group in benchmark categorisation tasks and outperformed several external large language models, including Claude Sonnet and Claude Opus.
The work centres on a private model built in the UK using NHS patient data, with hosting and governance kept within the UK. Patient data processed through the model is stored, hosted and governed under UK law, while model weights, fine-tuning, hosting and inference also remain in the UK.
The pilot used data from Patchs, OneAdvanced’s triage and online consultation platform, which supports more than 500,000 monthly patient interactions in the UK. The training set was a stratified, balanced collection of pseudonymised NHS patient triage requests submitted through the platform.
The project comes as NHS organisations face growing pressure to improve access to care while reducing clinicians’ administrative burden. It also arrives amid broader debate over data governance, privacy and whether sensitive public sector AI systems should rely on overseas infrastructure and models.
Benchmark tests
The pilot compared the model’s accuracy with a range of external large language models. OneAdvanced said the NVIDIA Nemotron-Nano-9B Care Navigation pilot delivered inference costs up to 150 times lower than leading frontier models.
Lower operating costs could help NHS bodies procure AI tools for use inside clinical workflows rather than as stand-alone technology projects, the company argued. It also said the model could reduce wasted resources by improving the accuracy of routing patients to the most appropriate care pathway.
Unlike public large language models, the system is designed to learn from clinician corrections over time. When GPs using the system amend its outputs, those changes can become part of the training data used to improve the model.
That feedback loop is central to the company’s argument that a healthcare-specific model trained on UK data can perform better in NHS settings than more general-purpose systems. OneAdvanced pointed to its long history in NHS software and access to large healthcare workflow datasets as the foundation for that approach.
Across its wider healthcare business, OneAdvanced said its software supports more than 40 million NHS patients annually. Its systems are used by 85% of NHS 111 services, more than 4,000 GP practices and over 160 NHS trusts.
Clinical use
Care navigation tools are used to identify the likely subject of a patient request, gather follow-up information and direct the patient to an appropriate response. In general practice, that can affect how quickly requests are reviewed and whether patients are directed to self-care, pharmacy, routine GP review or more urgent clinical attention.
Dr Ben Brown gave an example of how the model is intended to support that process in day-to-day care. “Detecting clinical topics in patient requests is key to enable the ‘AI Care Navigator’ within the triage and online consultation platform to ask the right follow-up questions and provide the right response to patients. This allows me, as a busy GP, to understand patients’ needs more quickly and deal with their requests more effectively. Increasing the accuracy of clinical topic detection through AI Care Navigator makes this process even better, reducing workload for me and my staff and ensuring patients get the most appropriate care more quickly,” said Brown.
Simon Walsh, chief executive officer of OneAdvanced, set out the company’s position on specialist healthcare AI. “We’re delighted to have invested in this technological advancement, delivering improved care navigation accuracy for patients and waste avoidance for the NHS. We are focused on delivering structural advantages to the UK NHS, along with compliance to medical device and UK sovereign AI model standards, utilising decades of experience and deep integrations with the NHS, which are an absolute necessity for real-world impact to be realised. The future of AI in healthcare will not be built on generic models. In healthcare, accuracy, governance and context matter more than anything,” said Walsh.
He added: “With decades of experience supporting the NHS and one of the UK’s largest healthcare workflow datasets, OneAdvanced is helping to transform how primary and community care best serve patients for the years ahead.”
NVIDIA framed the pilot as an example of sovereign AI in a regulated sector. “The UK has a significant opportunity to lead in the development of sovereign AI for highly regulated and mission-critical sectors such as healthcare. OneAdvanced’s work demonstrates how NVIDIA technology can be applied to create practical AI systems that support clinicians, improve patient experience and operate safely at scale. This pilot highlights the difference organisations with deep NHS expertise and context can make in accelerating digital transformation across healthcare,” said Hills.
