On 6 October, the UK government accepted all 44 recommendations of its commission on regulating AI in healthcare. Its plan includes stronger monitoring after products enter use and clearer responsibilities across the health system. An implementation roadmap is due by spring 2027. This is a commitment to reform, not proof of safer care. Its most consequential idea deserves attention beyond Britain: buying medical AI should create a continuing obligation to patients, with money and authority attached.

Imagine a district hospital purchasing software that helps prioritise scans. The demonstration is impressive; the licence seems affordable. Six months later, equipment changes, the software is updated, or a different mix of patients arrives. Who checks whether it still works well? Who can suspend it? Who explains a mistake to the patient?

These are hypothetical questions about an ordinary procurement decision. They expose the gap between acquiring a tool and sustaining a service.

The strongest argument for faster adoption is compelling. Waiting also has costs. The World Health Organization identifies potential benefits where medical specialists are scarce, including help interpreting images. A useful system could bring attention to a patient sooner and give a clinician support that was previously unavailable. Demanding endless preliminary evidence can protect existing inequalities as effectively as reckless deployment can deepen them.

But access to software is not the same as access to dependable care. WHO also stresses external validation, representative data and attention to a product throughout its use. A reassuring result elsewhere cannot settle how a tool will serve the people walking through this hospital’s door.

That continuing attention is work. Someone must investigate errors, assess changes, train staff and decide when to fall back on another procedure. If these duties simply land on a busy clinician, an apparent saving can become an unbudgeted burden. Responsibility without time or authority offers little protection to either doctor or patient.

The economics are uncomfortable. In their FUTURE-AI consensus guideline, published in the BMJ, researchers explicitly warn that implementation costs could disadvantage smaller developers and health systems with fewer resources. They also argue that building safeguards early can avoid expensive changes later. These are expert recommendations, not a measured guarantee of savings.

My inference is that safe adoption could become a source of market concentration. Large hospital groups may spread evaluation costs across many sites. Smaller facilities may depend more heavily on a vendor’s assurances. Established suppliers could turn the ability to finance monitoring into an advantage that a useful newcomer struggles to match.

The answer should include shared, independent testing capacity. Public purchasers, hospital networks and research institutions could pool parts of the work while retaining checks in each local setting. That would help small providers buy informed judgment without purchasing an entire research department. It would also make changing suppliers more credible.

For India, this is a concrete investment choice. The ICMR’s 2023 ethical framework already addresses development, deployment, governance and informed consent across a broad set of stakeholders. The practical challenge is to give those concerns a budget. A procurement programme should finance evaluation under the equipment, staffing and patient conditions in which the system will operate.

For a patient using a local language, meaningful explanation may require a person rather than an English screen. For someone travelling far for treatment, an accessible complaint route matters only if it can produce an answer without another costly journey. These requirements belong in the service design.

Three futures are plausible. In a better one, useful AI expands clinical capacity while shared evaluation and paid monitoring make dependable care affordable beyond wealthy hospitals. Suppliers earn recurring revenue by supporting continuing performance.

In an uneven future, prestigious hospitals buy the whole service while smaller ones buy the licence. The same technology then reinforces differences in institutional capacity.

In a harmful future, nobody notices deteriorating performance quickly, clinicians become the default bearers of blame, and patients face institutions and suppliers pointing at each other. That outcome would reflect choices about responsibility, not an unavoidable property of AI.

The UK’s announcement leaves implementation to be demonstrated. Buyers elsewhere need not wait to improve contracts: name who monitors performance, who pays for investigations, who may stop the system and what happens when the supplier leaves.

Does the business still work after paying for continuing evaluation and support? Can a small hospital afford the complete service? Can staff challenge the tool without penalty, and can patients obtain a remedy? A credible medical AI business should be able to answer before the first patient depends on it.