·

·

12

min de lectura

Ambient voice technology (AVT): what it is and how it works

Ambient voice technology, or AVT, listens to a consultation and drafts the clinical note for you to review. How it works and what UK guidance says.

Doctor in middle of a consultation while ambient voice technology (AVT) listens and writes the clinical notes from the appointment in the background. Tandem AI Scribe is visible on computer screen behind the doctor.

Ambient voice technology (AVT) is AI-enabled software that listens to a consultation and drafts the note for you to check and approve. That check is where safety is decided: when seven products were tested on simulated UK primary care consultations, most errors were omissions, which a fluent draft does not show.

In short

  • Check each draft for errors or omissions: in UK testing, invented or inaccurate details were rarer than omissions but more often clinically serious.

  • A place on NHS England’s AVT supplier registry is the supplier’s own declaration. The clinical safety case, the data protection impact assessment and your own accuracy monitoring are your organisation’s.

  • In Great Britain, a scribe that only drafts notes for your review is not a medical device, but a feature that supports diagnosis or acts without your review makes it one.

  • Reported time savings come from observational, largely self-reported studies, and NHS results from an NIHR-funded evaluation are due in February 2027.

How ambient voice technology works

The software does the listening and the drafting. Telling the patient and checking the draft stay with you.

  1. Tell the patient at the start that you are using AVT, and respect it if they object.

  2. Start the software, which captures the conversation, transcribes it and drafts the note, letter or summary in your template.

  3. Read the draft against what happened, then correct and complete it.

  4. Approve it before anything is saved to the record.

The software drafts only from what is said aloud. A rash you look at or a pulse you count in silence reaches the note only if you say it or add it at review. In a Midlands pilot in primary and secondary care, clinicians said they had to voice their observations more explicitly so the system captured their assessment. Some felt this reinforced good consultation practice, while others found it disrupted how they naturally talked with patients.

Saying a finding aloud also tells the patient what you found. Usually that helps. A breast lump you plan to refer urgently deserves more care than a running commentary, so decide which findings to voice and which to type in at review. Microphone position matters too: NHS England’s guidance warns that distance and position can affect results, so try it in your own consulting room before the first clinic.

A worked example: one review, line by line

This consultation is invented to show the review step. No real patient is involved.

Joan, 71, comes to a blood pressure review with her son. Her amlodipine went up to 10 mg six weeks ago.

Joan: “My ankles have been swollen ever since the new dose.”

Her son: “She gets puffed out on the stairs as well. Mind you, so do I since my knee replacement.”

GP: “It’s most likely the tablet, but I’d like a blood test to rule out a heart problem. We’ll go back to 5 mg and add ramipril 2.5 mg for your blood pressure, with blood tests today and again in two weeks, and I’ll see you then.”

The GP physically examines both ankles and shins and takes Joan’s blood pressure, 148/86, without reading either finding out. The draft and her changes:

Section

Draft

What the GP changed, and why

History

Ankle swelling since amlodipine increased to 10 mg. Breathless on stairs. Recent knee replacement.

She kept the first two sentences and deleted the third, which is the son’s information and has no place in Joan’s record.

Examination

No section: the draft goes from the history straight to the assessment.

She added “Pitting oedema to mid-shin, both legs. BP 148/86.” She did both in silence, so the draft had nothing to work from.

Assessment

Heart failure.

She rewrote it as “Ankle oedema, likely amlodipine. Heart failure to be excluded.” The draft had turned a possibility into a diagnosis.

Suggested code

Heart failure.

She rejected it and coded the ankle oedema. Accepted, the code could have put Joan on the practice’s heart failure register and into its recalls, whatever the corrected text said.

Plan

Amlodipine reduced to 5 mg. Ramipril 2.5 mg started. Blood tests today and in two weeks. Review in two weeks.

She kept the medication changes and the review, and named the tests: NT-proBNP, kidney function and potassium today, kidney function and potassium again at two weeks. Aloud she had said only “blood tests”.

The lines that needed most work read as fluently as the right ones. “Heart failure” looks like a confident assessment, and the son’s operation sits naturally in Joan’s history. The missing examination leaves no visible gap, as the draft moves straight from history to assessment. A quick read for obvious mistakes and the GP edits, and approves this note.

Is an ambient scribe the same as dictation?

NHS England groups these products as ambient scribing products, sometimes called ambient scribes or AI scribes, and treats AVT as a major subgroup. It describes speech recognition and digital dictation as earlier versions of the same category. The difference that matters at work is who composes the note. With dictation, you speak the note and the software types your words. With AVT, the software hears the whole consultation and writes the note for you to check. So the errors differ: dictation mishears your words, while an AVT draft can leave out what you would have written and add what you would not.

Where AVT is used, and where to start

AVT is used across NHS care, from general practice to emergency departments. In a survey of 598 UK GPs in September and October 2025, 40 per cent were using an AI scribe and another 23 per cent had used one before. Concerns about safety and medicolegal risk were common, particularly among GPs not using one. Of those who had stopped, the commonest reasons were lack of practice support (32 per cent) and medicolegal and liability issues (28 per cent).

Some consultations give a draft that is easier to check than others. Begin with those, and add the harder ones once you know how the product handles them.

Consultation

What to do

Why

A routine one-to-one appointment in English

Start here.

Two voices and a familiar structure make the draft quickest to check against your own memory of the consultation.

A relative, carer or interpreter in the room

Check who the draft says said what.

The software has to tell the voices apart, and another person’s words can land in the patient’s history, as the son’s operation did in Joan’s.

A consultation the product has translated

Arrange an extra check by someone who understands the language.

NHS England’s IG guidance for professionals says translated consultations may need additional checks.

A strong accent or dialect, or a speech impairment

Read the draft line by line.

NHS England warns accuracy may vary. One UK simulation of a single product found no significant difference across accents, but phonological impairment markedly reduced recognition.

A highly sensitive conversation

Consider not using it at all.

The IG guidance names highly sensitive conversations as a case where an ambient scribe may not be appropriate.

Is AVT allowed in the NHS, and who checks it?

AVT can be used in NHS care. In England the assurance work and the decision to adopt sit with the organisation that uses it, which in general practice means the practice.

Who

What they say

What it means for you

NHS England, adoption guidance(version 3, updated 29 July 2026)

Assign a Clinical Safety Officer, complete DCB0160 and a Data Protection Impact Assessment (DPIA), check the supplier’s DCB0129, and comply with the Data Security and Protection Toolkit (DSPT) and Cyber Essentials. Pilots can continue only if they are time-limited, up to four months, and do not bypass compliance.

In a practice someone has to take the Clinical Safety Officer role: a partner, or a CSO shared across your primary care network. You need not do it alone: the guidance points anyone without the tools or capability to their local ICB, expects the ICB to be involved before any pilot, and says example templates are in development.

NHS England, adoption guidance

Users review and approve outputs before anything else happens, and monitoring data on accuracy is collected and analysed independently of the manufacturer.

A supplier’s accuracy figures do not replace your own.

NHS England AVT supplier registry

All assurance and decision-making sits with your local NHS organisation.

A listed supplier still goes through your checks.

MHRA guidance, 29 July 2026 (Great Britain: England, Wales and Scotland)

Products intended solely for transcription, summarising, drafting letters or suggesting clinical codes for a clinician to review are not regulated as medical devices. Products that support diagnosis, treatment or prevention, or that take automated action such as placing orders without clinician review, are regulated as medical devices.

The features you switch on decide which side of the line your product sits.

CQC, as set out in NHS England’s adoption guidance

Inspection teams may check whether the technology is safely deployed and whether staff have been trained to use it.

Keep a record of who was trained, and when.

HSSIB

An investigation into AVT in acute adult secondary care was notified on 6 August 2026, with a report expected in summer 2027.

There are no findings yet, but its notice reports that adoption is accelerating while the safety implications are not fully understood.

For a clinical safety officer, the MHRA row is the one to revisit after go-live. A product can cross the line with an update. The MHRA expects manufacturers to reassess the device rules with every change, so a release that adds suggested diagnoses or automatic orders is a question for your hazard log too.

What the evidence shows so far

Studies report shorter consultations and less time on notes, but the designs are observational and most of the time figures are self-reported. A UK rapid systematic scoping review of 21 studies, funded by the NIHR, concluded in July 2026 that current evidence does not yet support widespread adoption of AVT. It also found evidence on patient safety and cost sparse. The NIHR is also funding an evaluation in four NHS trusts, covering mental health outpatient services, acute outpatient clinics and A&E departments, with results expected in February 2027.

Study

What it found

Weakness

Seven AI scribes tested in UK primary care, 2025

Accuracy was high overall, but no product produced an error-free summary of eight standardised consultations. Omissions made up 83.8% of errors, a median of one to six per consultation depending on the product, and hallucinations, though rarer, were more often clinically serious.

The consultations were simulated and the products anonymised.

Danish general practice, 2026

20 GPs reported shorter documentation time after AI-assisted documentation began.

The time was self-reported in categories, and the authors call for controlled studies with objective time measurement.

Capio Ramsay Santé, Sweden, 2026

The study analysed 236,153 notes from 1,295 clinicians. Among 177 fully onboarded clinicians, estimated documentation time per note fell from 6.69 to 4.72 minutes.

It evaluated Tandem’s own scribe, and Tandem co-authored it: two authors are Tandem employees and one is its chief executive. The time figure was self-reported in a single-arm design.

Emergency care in 48 Spanish hospitals, 2026

Over 12 months, emergency consultations that used an ambient scribe were 21.8% shorter on average.

It was retrospective and observational, its authors work for the hospital group that deployed the scribe, and it measures consultation length rather than documentation time.

Systematic review of AI voice-to-text documentation, 2025

Safety findings were inconclusive: three of six studies that examined safety raised concerns.

Seven of its nine studies were from the US.

The time studies say little about what consulting with a scribe is like. In the Midlands evaluation, automated note-taking often improved clinicians’ interaction with patients. But outputs were sometimes generic, reinforcing defensive documentation and reducing personal tone. A few users found that a summary read later did not bring the consultation back to mind the way their own notes did.

Who is responsible for an AVT note?

You are professionally responsible for what AVT records on your behalf. NHS England’s IG guidance makes the user accountable for the accuracy of what is added to the record. It asks you to check the output every time, within a timeframe that allows a meaningful review.

In practice that means before the next patient where you can. The review takes time out of whatever the scribe saved, so count it when you judge the savings. On a list running late it is the easiest step to put off. By the end of a session, though, consultations blur, and a plausible draft is hard to check against a memory that has faded.

Read first for what is missing: the authors of the seven-product study advise checking for omitted psychosocial details and medications. Then read for what sounds right but is not, such as the plausible-sounding insertions they single out. NHS England’s adoption guidance warns of a hypothetical statement recorded as a confirmed diagnosis. Its example is a patient’s hypothetical, and Joan’s draft did the same with the GP’s “rule out a heart problem”. Suggested codes need the same care, as a wrong one outlasts the corrected text.

It is good practice to show that a note or letter was written with the help of an ambient scribe, for example with a closing line on a letter. Information about another person that the record does not need, like the son’s operation, comes out at the same review.

Monitoring accuracy during a pilot

Your own review catches errors in single notes but cannot show how the product performs, because you check against your own memory. NHS England’s adoption guidance asks for accuracy data collected and analysed independently of the manufacturer, and for ongoing audits of clinical documentation. Agree a protocol with your clinical safety officer, who may well be you, before the first session, and run it through the pilot, which the guidance caps at four months.

  • Who checks. Someone other than the consulting clinician reads a fixed weekly sample of approved notes against the transcript, where the product keeps one for audit. Your clinical safety officer sets the sample size.

  • What they record. Each error as an omission, an inaccuracy or a hallucination, the seven-product study’s categories, graded by the harm it could have caused. Grade as well as count: in that study the rarer kinds were more often clinically serious.

  • When to pause. Criteria agreed in advance, such as an error that reached a record and could have caused harm.

  • What the partners see at the end. The sample results, the incidents and the hazard log, so the decision to continue rests on your own evidence.

Report through your local incident management system any error that reached the record and any near miss you caught that could have caused harm, and log each as a hazard. In England, that system feeds the national Learn from Patient Safety Events service. If the product is a medical device, report concerns to the MHRA through Yellow Card as well. HSSIB’s notice reports indications that routes for recognising and reporting AI-related incidents are not yet mature, so write that threshold into the protocol before you start.

Keep a smaller sample running after the pilot. Close reading gets harder once drafts usually come back right, and the adoption guidance lists that over-reliance, or automation bias, as a risk.

Telling the patient

The same IG guidance says you must tell patients at the start of any session that you are using AVT, which in general practice means each consultation. Its guidance for patients tells them to expect this, and that they can ask for it not to be used. If they object, respect it and explain how that might affect their care. Your organisation may have its own procedure for recording objections. The professional guidance includes example wording you can adapt, and that wording tells the patient the tool makes no decisions about their care.

Say it before the history starts. A patient who realises halfway through that the conversation is being captured has had no real chance to object. That is harder to put right than one sentence at the start.

What to check before choosing an AVT product

The first thing to check is how the product will connect to your clinical system. NHS England describes at least two routes: AVT from your existing EPR (medical record system) supplier, or an add-on from another supplier. In most general practice settings it expects integration with the principal EPR.

In the 2025 GP survey, most respondents worked on EMIS Web or SystmOne. Over half of current users had their scribe integrated into the clinical system, against just over a quarter of those who had stopped. So the question for each supplier is what happens in your system: does the draft land as free text or in structured fields, and does it suggest codes?

The supplier’s DCB0129 clinical safety case report belongs in your file, with its evidence on accuracy, including error and omission rates validated with clinicians. How those rates were measured matters: a figure from scripted scenarios tells you less than one from clinics like yours. Your DPIA also needs the supplier’s answers on data: how long the audio, transcripts and final outputs are each kept, and where data is accessed, shared and stored. Get those answers in writing before you sign.

On liability, a practice stands where a trust does. NHS England’s guidance says the non-delegable duty of care stays with the trust or primary care provider. Liability can default to it where a supplier lacks cover or does not fully accept responsibility. Clear contracts setting out roles, responsibilities and liability can limit that exposure.

n England, clinical negligence claims about NHS general practice for incidents since 1 April 2019 are covered by NHS Resolution’s Clinical Negligence Scheme for General Practice. Few practices have a legal team to involve, so before you start, ask your medical defence organisation how it treats notes drafted with AVT.

Where Tandem fits

Tandem makes an AI medical scribe that uses ambient voice technology, and if your practice uses Accurx, Tandem is the technology behind Accurx Scribe, powered by Tandem. The draft is yours to check, and nothing enters the medical record without clinician approval.

Tandem’s own commitment is that patient data is never used to train AI models. Treat it as you would any supplier’s data claim and check it through your DPIA, starting from how Tandem handles patient data.

Tandem’s AI Scribe is CE marked under EU MDR as a Class IIa medical device. EU MDR is the European Union’s Medical Device Regulation. In Great Britain the MHRA’s test, set out in the table above, looks at each of a product’s functions in the light of its intended purpose. So ask Tandem, as you would any supplier, which features you will use and how each is registered here, and record the answers in your DCB0160 and hazard log.
Tandem is also on NHS England’s AVT supplier registry, which is a self-certified listing. Its assurance is a DCB0129 safety case, the Digital Technology Assessment Criteria (DTAC), the DSPT and Cyber Essentials Plus: ask for each one for your DCB0160 file.

Explore Tandem’s AI medical scribe.


Frequently asked questions

▶ Which AI scribes has the NHS approved?

The NHS has no national approval for AI scribes, so none holds one. Since 16 January 2026, NHS England has run a self-certified supplier registry, expecting Digital Technology Assessment Criteria (DTAC) completion and MHRA registration where applicable. It endorses no listed supplier and runs only preliminary completion checks, and the registry is not a commercial framework.

▶ How much does ambient AI cost? ​

Each supplier sets its own price. The registry is not a commercial framework: each NHS body buys AVT through its own procurement. Ask how the price is structured (per clinician, practice or site) and how sessional and locum GPs are counted. Ask too what integration adds, which NHS England warns can be substantial, and whether costs rise with stored data. Before buying, ask your ICB whether it funds or commissions one centrally.

▶ Do patients have to agree to ambient voice technology?

They must be told at the start of the session, and if they object you respect it. Explicit consent is not needed when an ambient scribe supports individual care in England. NHS England’s IG guidance says you can rely on implied consent under the common law duty of confidentiality. If a patient lacks capacity, check whether someone has legal authority to decide for them, or whether they objected before. Otherwise, decide in their best interest.

▶ Can ambient voice technology include clinical decision support?

Yes, and a feature that supports diagnosis or treatment makes it a medical device. The MHRA’s example is a feature estimating the probability of a condition, relied on to decide treatment. It says this may be considered decisive information for a diagnosis, which should be regulated as Class IIa. A “not for diagnosis” disclaimer does not change this if medical claims are made elsewhere.

Sources

Guidance

Regulation and indemnity

Investigation

Studies

Empieza a usar Tandem hoy

Únete a miles de facultativos que disfrutan de una documentación sin estrés.

Empieza a usar Tandem hoy

Únete a miles de facultativos que disfrutan de una documentación sin estrés.

Empieza a usar Tandem hoy

Únete a miles de facultativos que disfrutan de una documentación sin estrés.