.png)
.png)
Somewhere in most NHS organisations, a member of staff is reading a discharge summary on one screen and retyping its medication list into another. The task takes minutes and recurs hundreds of times a week, yet it appears on no productivity dashboard. On 16 November OpenAI announced software that can do this work. Its new healthcare agents operate software much as a person would, reading the screen and typing into it, and they need no interface between the systems to carry information from one to the other.
The launch is aimed at the American market. Free clinician accounts are limited to verified US clinicians, and the compliance language is that of HIPAA. CVS Health has used the agents to process referrals and move documents between internal systems, and OpenAI says one cleared a months-long backlog over a weekend. Google launched a rival product six days earlier, in US preview, with connectors reaching more than 400 healthcare applications. Neither company is addressing Britain. The relevance to the NHS is emerging and contingent, but the problem these products target is one England knows well.
England's health service runs on thousands of separately procured systems, and the interoperability promised by successive national programmes has arrived unevenly. The 10 Year Health Plan commits to a Single Patient Record joining GP, hospital and social care data, with an ambition that patients can see it through the NHS App from 2028. NHS England is investing £10bn in technology over three years and has said it will favour AI tools integrated with the electronic patient record over standalone ones. The preference is telling. Integration is what makes these tools work, and integration is what much of the estate currently lacks. Until the record exists, trusts will carry large volumes of work that sits between systems unable to speak to each other.
Screen-reading agents offer to do that work without waiting. No supplier has to build an interface and no national programme has to deliver one. For a finance director facing productivity targets, and for staff weary of copying data between windows, the appeal is obvious. It will be pressed by vendors and will find a receptive audience wherever efficiency targets bite.
The evidence for reliability is thinner than the launch suggests. A Stanford benchmark published in April tested agents on insurance-related administrative work across a simulated records system, two payer portals and a fax queue. The strongest agent completed 36.3 per cent of full workflows end to end, although the best score on individual steps reached 82.8 per cent. Each workflow contained around a dozen steps, and a single error counted as failure. The models tested pre-date this launch, and OpenAI has not, according to reporting on it, disclosed workflow-level error rates for its new agents. A referral that is right nine times in ten is a manageable risk. A referral that fails silently is a clinical incident waiting for a date.
Governance is where the NHS would feel the difference. The clinical safety standards that apply to health IT, DCB0129 and DCB0160, assume a defined product with a manufacturer's hazard log and a deploying organisation's safety officer. An agent that navigates a screen is a moving target, because a supplier's interface update can change its behaviour without anyone informing the trust. Identity raises a second problem. If an agent acts under a clinician's login, the audit trail records a person who took no part in the action, which complicates both a Data Security and Protection Toolkit return and any later inquiry into a referral sent to the wrong service. Somebody will be accountable for those errors, and at present nobody has decided who.
Procurement adds a further layer. Capabilities of this kind rarely arrive as a discrete purchase. They appear inside platforms already bought, as the rollout of an AI assistant to more than half a million NHS staff has shown, and they place a model owned by an American company between NHS staff and NHS patient data.
Practical steps follow. Trusts should insist that any agent works under its own identity with its own log, and should test end-to-end success locally before scaling anything. NHS England and the Department of Health and Social Care should extend the clinical safety standards to cover agents that operate interfaces. Policymakers should resist treating agent capability as a reason to relax interoperability requirements on suppliers, since an agent that clicks through a poorly designed system leaves the design as poor as it was. Suppliers of electronic patient records will need to say whether they permit their software to be operated this way.
The NHS has a long record of temporary bridges becoming permanent roads. Screen-reading agents are the most capable bridge yet offered, and a Single Patient Record planned for 2028 gives the best reason to keep them temporary. If the agents deliver, the temptation will be to stop asking for the record. The test over the next two years is whether the service uses the agents to survive until the record exists, with failure rates published and interfaces still mandated, or lets them become the reason it never arrives.