The ambient scribe claim ladder: five rungs, five kinds of proof

The ambient scribe claim ladder: five rungs, five kinds of proof#

Foxes read a trail by its age as much as its direction, and a good claim deserves the same reading. In one week, one ambient scribe vendor has been in two headlines. On 28 September Heidi unveiled Heidi II, a version of its platform that acts on the administrative work around each visit rather than only documenting the visit, and Fierce Healthcare reported it on 30 September. The same day, Healthcare World told its readers that an AI scribe had released 11,000 clinical hours at Walsall.

Dating first, because dates are the commonest thicket in this quarry. The Walsall figures are not from this week. Walsall Healthcare NHS Trust published them on 17 October 2025, and Heidi's own case study carried them from July 2026. The Healthcare World piece is a fresh retelling of year-old trust numbers. That is not a crime, but a buyer who reads 'this week' into it is counting the wrong rings on the tree.

The sharper lesson is that the two stories make different kinds of claim, and a page that blurs them will be read as the boldest one. Every ambient scribe page has the same temptation. We find it useful to treat every ambient scribe claim as a rung on a ladder. Each rung needs its own proof, its own wording and, as the regulator's examples show later, sometimes its own legal status.

This piece is written for ambient scribe vendors, for the marketers who write their pages, and for the NHS trusts and practice groups who have to decide which sentences to believe. It is judgement, not legal advice, and where a figure belongs to a vendor or a trust we say so.

Ambient voice technology: where the regulator draws the line#

On 29 July 2026 the MHRA published guidance on ambient voice technology-enabled products, developed with NHS England and aimed at medical devices placed on the Great Britain market. HTN reported that products intended solely for transcription, summarising clinical conversations, drafting letters or suggesting clinical codes for a clinician to review are not regulated as medical devices under the current framework. Products intended to support diagnosis, treatment or prevention, or to take automated action such as placing orders without clinician review, are.

That sounds like a tidy hedgerow for any ambient scribe, but the useful part is in the nine worked examples inside the PDF. We have grouped them below. Notice that function alone does not decide the verdict. Claims do.

Example 7 is the one for marketers to pin above their desks. It describes an ambient scribe that solely provides a summary of a conversation, identical in function to a non-device, whose maker claims it guides diagnosis and treatment planning and improves patient outcomes. The MHRA's verdict is that it is a medical device, because the manufacturer is marketing it for a medical purpose. A rung-four claim on a rung-one product changes its legal status.

The guidance is explicit about where intended purpose lives. It is set by the claims in the instructions, labelling and promotional materials, including websites and adverts, and a general disclaimer such as 'not for diagnosis' does not rescue a product whose marketing implies a medical purpose. It also warns that a product launched without a medical purpose can acquire one as features are added, so manufacturers should reassess with every change. The MHRA guidance PDF reads like a letter addressed to the web team.

The buyer's side of the same picture is NHS England guidance on AI-enabled ambient scribing, last updated on 29 July 2026 and written for England. It tells deploying organisations to assign a Clinical Safety Officer, complete DCB0160 documentation and a data protection impact assessment, and to ensure users review and approve outputs. It also points buyers to the NHS England supplier registry, which HTN says listed 19 self-certified suppliers including Heidi Health. Ambient clinical documentation, the plain job of turning talk into a draft note, sits on the comfortable side of the line, but only while the copy keeps it there.

If you want the wider FDA and MHRA picture on generative AI claims, our earlier piece on AI medical devices and the claim a regulator can quote back at you covers it, so we will not repeat it here.

Agentic AI healthcare: Heidi II and the rung that acts#

The Heidi II launch is a clean specimen of agentic AI healthcare marketing, so it repays a slow prowl. According to the company's press release, agents can take on chart preparation, referrals, recalls and follow-up coordination, can be saved as routines, and run under approval controls configured by the clinician or organisation. The release says that agents in healthcare need supervision and should be met with friction, and that clinical judgement remains with the clinician.

• 2.8m

  patient visits a week supported, as Heidi reports

• 10m

  follow-up tasks a week surfaced, per Heidi

• 175m

  visits supported by the scribe to date, per Heidi

• 400m

  follow-on tasks surfaced since February 2024, per Heidi

Those are vendor figures, reported by Fierce Healthcare and the release, and they describe volume of use rather than benefit. A visit supported is not a visit improved, which is exactly the gap between our ladder's first rung and its higher ones.

Availability matters just as much. Heidi II began rolling out on 29 September in English and French, and the release states that these product capabilities will not be available in the UK or EU, as Fierce also notes. It adds that Heidi II's capabilities are staged by region and setting. For a UK buyer reading the Walsall story, the practical reading is that the trust's results concern the ambient scribe, not the agent layer, and a page that lets those two blur together is promising something the product is not currently offered to do in that market.

That is Heidi's co-founder, Tom Kelly, on X, and it shows the register a vendor can slip into once a product can act. 'Wild' is a lovely word in a founder's feed and a tricky one on a product page. We quote it as a vendor voice only, not as evidence of anything beyond enthusiasm.

Now put the MHRA's Example 8 beside the Heidi II description. An agent that places follow-up orders without clinician input or confirmation is a device. An agent that preps a chart, drafts a referral and waits for approval sits much nearer Examples 3 and 5. Heidi's own wording stresses approval, which points to the careful side. The marketing job for any ambient scribe that grows agents is to keep the verbs honest: 'drafts', 'prepares' and 'queues for approval' are three different promises from 'orders', 'decides' and 'completes'.

EU-bound vendors face a second clock. The Commission says it began enforcing the AI Act on 2 August 2026, with transparency rules for chatbots and AI-generated content starting the same day. Specculo, a regulatory consultancy, reports that the Digital Omnibus on AI fixes 2 August 2028 for high-risk AI in regulated products such as devices, and that the Commission's transparency guidelines list AI-generated summaries of text as content needing marking while transcriptions are excepted. That is the consultancy's reading of the Commission's documents, so treat it as a lead to the primary text.

Ambient AI scribe evidence: what a trust report can and cannot prove#

Rung two is the time claim, and Walsall is the best-documented ambient scribe example in this week's coverage. The trust reported that emergency department documentation fell by 62 per cent, from a range of 15 to 20 minutes to a range of 5 to 7 minutes, according to Heidi's case study, and that dermatology letters went from 9.28 days to 18 minutes. Those are the trust's reported figures, and a trust is entitled to be proud of them.

A careful buyer will still ask three questions. First, the headline hours: with around 300 patients attending daily and a saving the trust gives as six minutes a consultation, our arithmetic is 300 patients times six minutes times 365 days, or about 10,950 hours, which matches 'nearly 11,000'. Healthcare World's account says the trust estimates the time saved 'could' release nearly 11,000 hours a year, which is the language of a projection. It assumes every attendee is documented with the scribe all year, not a count of hours actually released.

Second, the numbers do not sit perfectly together. A 62 per cent cut from a 15 to 20 minute baseline implies a saving of nine to twelve minutes, while the trust's release speaks of six minutes on average. Both come from the same trust announcement, and the likeliest explanation is different measures or samples, which is exactly what a buyer should ask about. Third, the trust's later August 2026 announcement describes an eight-month pilot across six specialties, an average documentation time of less than four minutes per patient, and Heidi as preferred supplier for two trusts. It is a newer and softer-edged ambient scribe claim, and it needs the same questions.

Rung three, wellbeing, is where controlled trials come in. A randomised trial published in NEJM AI in November 2025 assigned 238 outpatient physicians across 14 specialties to Microsoft's DAX Copilot, to Nabla or to usual care. Nabla users cut time in note by 9.5 per cent, while DAX users showed no significant change. Burnout, task load and work exhaustion measures suggested improvement for both, though the authors say those secondary findings need confirming in larger multicentre trials, and they note occasional clinically significant inaccuracies.

The trial's own products were not Heidi, and that matters for copy. Borrowed evidence is a common slip in this market: a vendor cites a trial of a category and lets the reader assume it tested the vendor. A defensible page names the product tested, or says plainly that the trial concerned other ambient AI scribe products and describes what a local audit has measured instead.

Accuracy is the shadow on every rung of an ambient scribe's ladder. A preprint posted to arXiv on 31 August 2026 audited three commercial AI scribes, which the abstract does not name, on 142 consultations. Its authors report that 31.3 per cent of notes carried a verified failure, and that the share ranged from 28 to 97 per cent depending on the standard applied. It is a preprint, not peer reviewed, and it uses model panels to judge errors, so it is a caution rather than a verdict. We chart it for the shape of the problem: the answer depends on the instrument.

What should a buyer demand from the vendor? NHS England says its supplier registry requires evidence of real-world benefit in the proposed care setting, and it asks deploying organisations to collect monitoring data independently of the manufacturer, to test for accents and dialects, and to remember that they may still carry liability. A good ambient scribe page, built on honest ambient clinical documentation, anticipates those checks before the procurement team ask.

Rung four, outcomes, is thinner for any ambient scribe. None of the sources we read for this piece measures patient outcomes for an ambient scribe, so a page that claims better care is claiming ahead of what we could verify, and under the MHRA's Example 7 it may also be claiming itself into the device rulebook.

How to word each ambient scribe rung without writing a device#

Here is the working version, offered as judgement. For an ambient scribe, pick the rung your evidence has really reached, write the verb that rung allows, and name the evidence beside it.

Three habits keep an ambient scribe page honest. Date every number, so a year-old trust result never poses as news. Name the product in any trial you cite. And write the approval step into the sentence, not the footnote. Buyers also search a vendor by name before they trust it, and what search engines say about that name is its own question, which our piece on branded keywords and AI Overviews takes up.

The same discipline shows up across regulated health marketing. The lessons in our posts on telemedicine certification and the CMS ACCESS model rhyme with this one: the claim is only as strong as the paper trail behind it. If you would like help building a claims register, our healthcare marketing team, our content marketing services and our AI consultancy can take it from audit to page, and our SEO and GEO services look after how AI engines then quote it.

Frequently asked questions#