For more than two months, an intensive care nurse at Barts Health watched a patient with a tracheostomy struggle to communicate. Unable to speak, the patient had to find other ways of making their needs understood, while the nurse witnessed first-hand the frustration and isolation that can accompany such a loss of communication. That experience has since developed into a software project with potentially far-reaching implications. Anna Lazcano's Project VOCAL, short for Voice for the Critically Unwell, is designed to give patients a way to communicate by selecting words and phrases through eye movements or touch, with artificial intelligence helping to assemble those selections into coherent messages. The project remains in development, and its prototype is currently being reviewed by speech and language therapists and AI specialists. Lazcano has made clear that a working version will be deployed only when experts consider it appropriate. In a healthcare system increasingly keen to demonstrate its technological ambitions, that insistence on clinical scrutiny deserves attention. It is precisely the kind of caution that should accompany every AI project entering patient care.

VOCAL is part of a wider programme of technological experimentation at Barts Health, where several applications are already moving beyond the drawing board. The trust says it was the first NHS trust to introduce an AI-powered telephone service, Ovia, which contacts patients on heart and lung virtual wards at agreed times, asks straightforward questions and relays their responses to clinical teams. Its introduction followed a practical problem with smartphone-based services: some patients could not use the applications and were consequently unable to participate in remote monitoring. The decision to offer a telephone alternative reflects an important reality of digital healthcare. Technology that assumes every patient owns a suitable smartphone, understands how to use an app or feels comfortable navigating digital services risks excluding precisely those who may need support most. Elsewhere, more than 2,000 clinicians across Barts now have access to Oracle's ambient voice technology, which transcribes consultations and prepares draft patient letters as part of a national trial. Seven hundred clinicians used the system in September alone. These initiatives have emerged from identifiable problems within clinical practice, rather than technology looking for somewhere to be deployed. That distinction matters when the NHS is under pressure to demonstrate tangible improvements from its investment in AI.

The Government is now attempting to give this activity a more national direction. The October Budget introduced an annual Chancellor's AI Challenge Prize, with the first round offering a top award of up to £2.5 million for solutions addressing priority healthcare challenges. It also commissioned an independent review of NHS productivity and the opportunities presented by technology, with findings expected early next year. The ambition is considerable, particularly at a time when ministers are looking to technology to improve productivity across an overstretched health service. Yet the NHS has never been short of promising ideas. Its persistent difficulty has been turning successful local experiments into dependable services that can operate across organisations with different clinical systems, procurement arrangements, staffing pressures and tolerances for risk. Recent developments illustrate the problem. Digital Health and Care Wales abandoned a £10 million ambient voice procurement this month after too few suppliers met its requirements. Meanwhile, the south east ambulance service, having recently surveyed its paramedics on AI, confirmed in a freedom of information response this year that it had no standalone AI strategy. Neither development should be treated as a scandal in isolation. Together, however, they expose a familiar weakness in NHS technology adoption: procurement and experimentation are advancing while the structures needed to guide them remain underdeveloped.

The views of the clinical workforce offer a clearer indication of what successful adoption will require. Research at South East Coast Ambulance Service, based on 171 survey responses and 14 interviews, found considerable support for using AI in clinical care, provided appropriate safeguards were established. Ninety-two per cent of respondents wanted paramedics involved in the design and testing of AI tools, while 96 per cent believed that final clinical responsibility should remain with the paramedic. Around 80 per cent said clear governance, audit and accountability arrangements would increase their confidence in the technology. The research is relatively modest in scale and reflects the experience of one ambulance trust, so its findings cannot automatically be applied across the NHS. Nevertheless, the responses offer a useful warning to policymakers and technology developers. Clinicians are not rejecting artificial intelligence. They are asking to retain control over decisions that affect patients and to have a meaningful say in the systems they are expected to use. That is a reasonable condition for adoption, particularly when the consequences of an error can extend well beyond an administrative inconvenience.

Meeting those expectations will require considerably more than publishing guidance or asking trusts to establish local oversight arrangements. Meaningful clinical involvement in the design of AI systems takes time, and that time must come from a workforce already managing substantial workloads and waiting-list pressures. Effective audit and accountability also require people with the expertise and capacity to examine how these systems perform once they enter routine practice. Trusts operating under tight financial constraints are unlikely to prioritise such work without clear national expectations and funding to support it. The Chancellor's prize may attract technology companies, health-tech developers and life sciences firms keen to demonstrate what their products can achieve. Yet winning recognition for an innovation is only the beginning of its journey into the NHS. Developers will eventually face procurement teams that need clear specifications, clinical leaders who require convincing evidence and organisations that must establish who is responsible when something goes wrong. Patients, meanwhile, may encounter AI through an automated telephone call or a letter drafted by software without necessarily understanding how much of the interaction has been handled by a machine. Their confidence will depend, in part, on being told openly when and how these tools are being used.

Project VOCAL brings these concerns into particularly sharp focus. Its intended users are patients who may be unable to speak, making the consequences of an inaccurate message unusually serious. Someone receiving intensive care and relying on eye movements to communicate may have little opportunity to challenge a sentence that the software has misunderstood or explain that the message it has produced does not reflect their wishes. In such circumstances, the usual safeguards of patient feedback become harder to rely upon. Independent clinical scrutiny, careful testing and ongoing monitoring must therefore carry much of the responsibility that a patient's own voice would otherwise provide. They cannot be treated as administrative steps to complete after deployment. The technology must earn the right to be used before it reaches the bedside. The Budget has put public money behind the pursuit of AI innovation. The independent review due early next year will offer a more revealing test of the Government's intentions: whether it is prepared to invest in the governance and scrutiny that will determine whether those innovations can be trusted.