In consulting rooms across England, a quiet piece of software has begun listening in. It sits through the appointment, turns the conversation into a structured note, and hands the clinician a draft summary before the patient has left the building. For many doctors this has become the unremarkable texture of a Tuesday morning clinic. For NHS England and the medicines regulator, it has become the subject of a pointed warning to those who run hospitals.

A joint letter from NHS England and the Medicines and Healthcare products Regulatory Agency has told trust boards they must take direct responsibility for the safety of these so-called AI scribe tools, setting out eighteen assurance items covering oversight, governance, staff training and procurement. Among them is an explicit requirement that clinicians review and approve every AI-generated output before it is used in patient care. The letter arrives as national technical standards for the tools, rather than being tightened, are being streamlined to speed up adoption.

The appeal of ambient scribing is straightforward. Administrative load has long been cited as a driver of burnout among doctors, and software that listens to a consultation and drafts the note, letter or referral in the background promises to return minutes, sometimes many minutes, to clinical time. That promise has already been translated into scale. One region recently completed what is understood to be the largest deployment of the technology so far, covering some seventy thousand clinicians across fifteen trusts and over a thousand GP practices, built on a national programme first hosted by a single acute trust in Lancashire.

What has changed is where the assurance now sits. NHS England has spent much of the past year building the scaffolding for safe adoption: a supplier registry, the Digital Technology Assessment Criteria, and the established clinical safety standards known as DCB0129 and DCB0160, which require manufacturers and deploying organisations respectively to document risk and maintain oversight. None of that scaffolding is being dismantled. But the letter makes plain that where a product is used, and how its outputs are checked, is a judgement for the trust to own, not a guarantee the centre can supply. NHS England's own guidance is explicit on the point, describing a non-delegable duty of care that remains with the trust or primary care provider regardless of which supplier's product is running in the background.

This distinction matters more than it might first appear. Trusts have already shown how easily good intentions can outrun proper process. One acute trust abandoned plans this spring to trial a major supplier's voice technology after concerns emerged that the product had not secured the accreditation NHS England requires, a decision that only came to light through trade press scrutiny rather than routine disclosure. The episode is a reminder that procurement decisions taken with genuine confidence in a tool's promise can still leave a board exposed if the paperwork behind it has not kept pace.

For chief executives, chief medical officers and chief information officers, the practical consequence is a lengthening list of things to get right before a scribe is switched on anywhere in a trust, and to keep getting right afterwards. Clinical safety officers must be in place, audit and monitoring must be continuous, and the habit of clinicians actually reading and correcting what the software produces, rather than signing it off unread, has to survive the pressure of a full clinic list. Regulators are trusting that discipline to hold at eight or nine hospitals rather than one committee room in Leeds.

The government's ambition for an NHS run substantially on artificial intelligence has not slowed. What this letter confirms is that the ambition and the accountability are travelling by different routes, and that when something goes wrong in a consultation room, it will be the trust board, not the software supplier, answering for it.