MHRA's AI scribe line, 5 board checks: this week in AI

If you have been sitting on an AI scribe decision because nobody could tell you whether the software counts as a regulated medical device, that question now has a published answer. The MHRA gave it on 29 July, and almost every write-up since has filed it under NHS news. It is not only NHS news. Medical device classification attaches to the product, not to the organisation buying it, which means the line the MHRA drew runs straight through independent practices too. That story leads this week, alongside two separate five-point AI governance lists published days apart on opposite sides of the Atlantic, a funded testing programme that will finally produce scribe accuracy data nobody is selling, and a survey number worth more to most practices than any of it.

1. The MHRA said which AI scribes are medical devices, and which are not

On 29 July 2026 the MHRA published guidance, developed with NHS England, setting out how existing medical device law applies to ambient voice technology products across Great Britain. Four functions sit outside the regime: transcribing a consultation, summarising a clinical conversation, drafting correspondence, and suggesting clinical codes for a clinician to review. A product intended to support diagnosis, treatment or prevention, or one that takes automated action such as placing an order or triggering a referral without clinician review, is a regulated medical device and must meet the associated safety and performance requirements. Source: MHRA press release and NHS England publication PRN02607_i, both 29 July 2026.

The number: four. Four named functions the MHRA has now confirmed do not, on their own, make a product a medical device under the Medical Devices Regulations 2002. The guidance changes no law. It states how the regulator reads the law it already has, which is precisely what suppliers and buyers had been asking for.

The so-what: read the sentence that decides the classification, because it is about intended purpose and automation, not about how clever the tool is. A scribe that drafts a letter for you to sign is outside the regime. A tool that reads the consultation and proposes a working diagnosis, or that fires a referral without anyone reading it, is inside it. Ask every supplier one question in writing: does any function of your product act before a clinician approves it? A supplier who cannot answer that cleanly has told you something useful about their own governance. This also sharpens the accuracy problem covered in the dental AI scribe hallucination rate no vendor agrees on, because falling outside device regulation means nobody is checking the accuracy claim for you.

2. NHS England published a five-point board checklist, and it is free to copy

The same day, NHS England wrote to trust chief executives, chief information officers and integrated care board leaders setting out what boards should assure themselves on before deploying ambient voice technology. The five areas: oversight and accountability, with a named human responsible for reviewing and approving AI output before it is relied on; local governance covering clinical safety and information governance; staff training on what the tool can and cannot do; risk management covering privacy, consent, information quality and automation bias; and commercial and procurement due diligence, including contractual rights to audit performance at AI model update milestones. NHS England also confirmed national implementation support beginning in outpatient settings and an AVT Centre of Excellence built with trusts that have already deployed. Source: NHS England publication PRN02607_i, 29 July 2026.

The number: five. Five assurance headings, written for organisations with boards, clinical safety officers and procurement teams, published under an open government licence.

The so-what: an independent practice has none of that machinery, and does not need it. What it needs is a document that answers the same five questions about its own tools, because those five headings map almost exactly onto what a CQC well-led assessment asks. Write one paragraph under each heading, name yourself or your clinical lead in the first one, and date it. That is your AI governance policy, and it took an afternoon rather than a consultant. If you want the longer version of what an inspector actually looks for, we set it out in the 10 AI governance checks CQC will look for in 2026.

3. Five NHS trusts are getting funded, vendor-neutral scribe testing

The National CLEAR Programme, hosted by East Lancashire Hospitals NHS Trust, has opened expressions of interest for CLEARvalidate, a fully funded six-month pilot supporting five NHS trusts to test the safety and performance of ambient voice technology across up to three clinical specialties each. It is designed as continuous safety assurance rather than a one-off report, and it is vendor-neutral. Acute, community and mental health trusts can apply until 21 August 2026. Source: National CLEAR Programme announcement, 13 July 2026, and Digital Health, July 2026.

The number: 21 August 2026. Twelve days from now, when applications close.

The so-what: no independent practice can apply, and that is not the point. The point is that the only credible AI scribe performance data anyone will be able to cite in six months is being generated right now, and none of it exists yet. Set that against what happened when a government auditor actually tested the category, covered in the government test that found problems in all 20 AI scribes. Until independent results land, treat every accuracy percentage in a sales deck as unverified, and negotiate a twelve-month break clause rather than a three-year term. The evidence base is about to move, and you want to be able to move with it.

4. 28% of patients have skipped cancelling because it meant ringing you

A patient survey published by Tebra in August 2026 found that 28% of patients had put off cancelling or rescheduling an appointment because the practice required a phone call rather than a text, a link, an app or a portal. Source: Tebra, "Where patients skip care", August 2026. Flag on that figure: Tebra sells practice management and patient communication software, so this is vendor research on a question its own products are sold to answer. Treat the direction as sound and the precise number as unaudited.

The number: 28%. Better understood as follows: roughly one patient in four who wanted to release a slot back to you decided the phone call was not worth it, so the slot went unfilled and was recorded as a no-show.

The so-what: this is the cheapest item on this week's list and mostly is not an AI problem at all. A cancellation you never receive is indistinguishable from a patient who simply did not turn up, and you cannot rebook a slot you do not know is free. Before evaluating anything ambient or agentic, check whether a patient can cancel and rebook without speaking to a human, and if they cannot, fix that first. The same reasoning applies to inbound calls that never reach you, which we worked through in why voicemail is not the safety net practices assume.

5. The first US health system was certified for responsible AI use

Hackensack Meridian Health announced on 29 July 2026 that it is the first US health system to earn the Joint Commission's Responsible Use of AI in Healthcare certification, a voluntary programme the accreditor launched earlier this year. Assessment covers five domains: governance and oversight, safeguards for patient privacy and health data, processes to identify and reduce risk and bias in AI tools, monitoring that tools continue to perform safely, and transparency and education. Source: Hackensack Meridian Health, 29 July 2026, and Healthcare Dive, July 2026.

The number: one. One certified organisation, from a US health system population in the thousands, several months after the certification became available. Voluntary AI accreditation is at the very start of its adoption curve.

The so-what: no independent practice is going to pursue a health-system accreditation, and none should. What is worth two minutes is the overlap. Put the Joint Commission's five domains beside NHS England's five board checks, written by different regulators in different health systems within days of each other, and four items appear on both lists: a named accountable human, staff training on the tool's limits, monitoring after go-live, and transparency with patients. When two independent bodies converge on the same four things, that is as close to a settled answer on "what does good look like" as this field currently offers.

What this week adds up to

Two useful things happened. The MHRA gave British practices a test they can actually apply to a product, and it takes one question to a supplier rather than a legal opinion: does anything here happen without a clinician approving it first? Separately, two regulators independently published governance expectations that agree on four points out of five. Between them you can build a one-page AI policy this month that would stand up to a CQC conversation, without buying anything.

The evidence gap is the part still unresolved. Regulatory clarity arrived before performance data did, which is an unusual order of events and slightly awkward for buyers. The MHRA has confirmed that most scribes are not medical devices, meaning nobody external is validating their accuracy, while the programme that will produce independent accuracy data has not started. Frankly, if you are choosing a scribe in the next quarter, the maths favours a short contract, a written answer on automation, and a documented sample check of the notes your clinicians sign, rather than waiting for evidence that is at least six months away.

Two dates for the diary. Colorado HB 26-1195 takes effect this Wednesday, 12 August, restricting what AI may do inside a therapy session; we covered what it actually says, including the notice duty that applies even to practices using no AI at all, in Colorado's AI therapy law and the provision the headlines skipped. CLEARvalidate expressions of interest close on 21 August.

If you want a clear read on where your own practice sits against these expectations, the AI Opportunity and Growth Assessment benchmarks you against the CARE Framework in two weeks. Or book a 20-minute discovery call and we will talk through what this week's guidance means for your specialty.