Britain’s National Health Service Needs an AI Maintenance Budget

Britain’s National Commission into the Regulation of AI in Healthcare published its recommendations on September 10, proposing oversight that follows medical AI into everyday use. The proposals include recording which software version affected a patient’s care and allocating safety responsibilities in contracts between manufacturers and healthcare providers. A government response is still to follow.

The consequential shift is from judging a product to supporting a service. For the NHS, Britain’s public health service, that raises a spending question: who will pay for the work after installation? My reading is that the commission’s approach can make AI more useful, but only if hospitals purchase the capacity to supervise it alongside the software. A contract can allocate a duty without creating the time to perform it.

The work behind the licence

Consider a hypothetical hospital using AI to help prepare clinical notes. Buying the tool is the visible transaction. Less visible work includes checking corrections, investigating recurring mistakes, training new staff and responding when an update changes the output. Those activities consume working hours even if the supplier includes technical support. The distinction matters because a hospital could save time on transcription while giving some of it back through verification and troubleshooting.

This is not an estimate of what any particular product costs. It is a way to distinguish gross savings from usable capacity. Minutes removed from one task do not automatically become minutes available for patient care. The relevant comparison includes the new tasks, who performs them and whether they fit into the working day. An adoption decision built around the licence fee alone leaves that comparison unfinished.

The commission proposes that manufacturers specify the conditions needed for safe use and that contracts explicitly assign responsibility for required risk controls. It also calls for better reporting and traceability after deployment. These are meaningful proposals because they bring work usually hidden behind the word implementation into the agreement between buyer and seller.

Demand is not readiness

There is evidence that clinicians want useful AI while doubting the systems around it. In January, the Royal College of Physicians reported that 70% of respondents to its member survey supported widespread NHS adoption, while 68% thought the necessary digital infrastructure was lacking. The survey was conducted in June 2025. It records respondents’ views, rather than providing a current technical audit of every hospital.

The college also called for clinicians to have time to help develop AI tools and for those tools to work with existing patient records. This independent professional perspective complicates the familiar account of reluctant staff holding back progress. In these findings, enthusiasm and concern coexist. Buying more software cannot, by itself, resolve a problem with the systems and working arrangements into which that software must fit.

The incentive problem runs in both directions. A supplier benefits when its product is easy to purchase; a hospital benefits when it is easy to operate safely. Those interests can align, but the purchase agreement has to connect them. If monitoring falls to the hospital, the hospital needs access to usable records and support. If the supplier owns an investigation, the hospital needs a clear way to raise it and learn what happened. Otherwise, each side can complete its assigned paperwork while the practical problem remains.

Make responsibility usable

Patient confidence adds another operational demand. In its response to the commission, Healthwatch called for transparency about AI use, human checks and patient choice about AI scribing, which turns consultations into draft notes. Explaining a tool, answering concerns and accommodating a patient’s preference are part of delivering it. Those interactions should appear in the service design, rather than being treated as friction to remove from an adoption target.

There is a serious objection to making every deployment carry an elaborate new monitoring apparatus. It could consume the savings, delay useful products and favour large suppliers that can afford more compliance work. The commission itself advocates oversight proportionate to risk and benefit, including ways to reduce reporting burdens. Its proposals should not be read as a demand for the same process around every application.

The practical answer is to make the workload visible and then reduce it. A hypothetical tool that records its version automatically could spare staff a manual step. A shared reporting route could replace duplicate forms. A supplier that makes corrections easier to inspect could reduce review effort. These are design choices worth rewarding in procurement because they improve the whole service, even when they do not change the model’s headline performance.

Before scaling a deployment, a hospital should be able to show where the claimed time goes: what is saved, what oversight adds and who carries the remaining work. Evidence that these costs are already measured, funded and falling would weaken the concern that supervision will become an unfunded duty. Without that evidence, a promise of shared responsibility is incomplete. The next stage of NHS AI adoption needs an operating budget that lasts beyond the launch.