MHRA has drawn the line on ambient voice technology.

The dust has settled on the debate around MHRA's new guidance on ambient voice technology and what does — and doesn't — constitute a medical device. But there is an important implication most Trusts are missing: what this means for AI in the NHS more broadly. The guidance itself is narrow and clear: MHRA's attention is fixed on the highest-risk ground, diagnosis and treatment. 

Everything below that line — the clinical and operational admin that consumes a third of clinical time — is open ground, and it is where the biggest gains sit: operational productivity and patient care.

The guidance is about Ambient Voice Technology (AVT). The reasoning isn’t. 

MHRA's guidance is written for ambient voice technology specifically. But it doesn't treat AVT as a special case — it says outright that AVT products are built on the same GenAI and large language models now running through every other AI tool in the health system. The test it applies — diagnosis, treatment, and clinical decision-making sit on one side of the line; administrative support sits on the other — is a test about what the AI is doing, not what kind of AI it is. Our understanding is that there is no obvious reason it stops applying the moment the input changes from speech to text. 

That's not just our reading of the guidance — it's where national policy is heading too. The National Commission into the Regulation of AI in Healthcare reached the same conclusion in its recommendations for a future regulatory framework: oversight should focus on the function that meets the medical-device threshold, not the product or the modality delivering it. It goes further, recommending the exemption be written into the medical device regulations themselves — not left as guidance MHRA can revise at will, a materially more durable line than the one this piece is built on. 

MHRA's own worked examples make the line unusually concrete.

What that opens up, in practice: 

• MDT preparation — pulling a patient's history together from multiple EPR systems ahead of a multi-disciplinary team meeting can take a clinician one to two days per case. Summarising and structuring that record for review is exactly the “administrative support” pattern MHRA has confirmed sits outside device regulation. 

• Referral letters — if discharge letters are administrative support with the clinician reviewing and signing off, the same logic extends to referral letters, drafted by AI and reviewed by a human before they go anywhere. 

• Population health case-finding — AI searching clinical notes for patients who may meet a clinician defined criteria for a known condition, surfacing names for a clinician to review. It is doing the notes search faster; the clinician still makes the call. 

• Discharge and flow bottlenecks — an operations manager using AI to cross-reference EPR discharge criteria against actual bed occupancy and flag patients stuck in the system for longer than expected, for a human to investigate and unblock. No clinical judgement is made by the tool — it is surfacing a process delay, not a treatment decision. 

What that means for how Trusts get AI capability: buy is no longer the only route. Where AI supports diagnosis or treatment, the calculus hasn't changed — for now: that's regulated territory, and buying a certified product remains the right answer, though the Commission's proposed staged authorisations are designed to loosen exactly this over time. But where it doesn't, Trusts already holding Microsoft 365 Copilot, or an equivalent, may have a viable route to develop selected use cases, subject to local assurance and the intended purpose of the tool, at close to zero incremental cost. 

Why the frontline has lagged in AI use-cases 

The caution wasn't unreasonable. Until now, working out whether an AI tool crossed into medical device territory needed specialist legal expertise to interpret the UK MDR — and getting it wrong carries real regulatory consequences. 

Faced with that ambiguity, Trusts have defaulted to caution over the ground that matters most: the clinical and operational frontline, where most of the workforce sits and where the Trust’s core purpose is actually delivered. 

That caution has shown up in one of two patterns: 

• Buy. Where Trusts have wanted AI in a clinical or clinical-adjacent workflow, the safe route has been to purchase software already certified as a regulated medical device — expensive, and dependent on a third-party vendor's roadmap and pricing for anything beyond what was bought. 

• Build, but only in the back office. Where Trusts have built AI capability internally, it has stayed almost entirely in finance, HR, and other core services — functions with no plausible read as clinical decision-making, and therefore the only ground that felt safe to experiment on without legal sign-off. 

Both are sensible responses to a genuinely unclear rule. But together, they have left the frontline — the majority of the workforce, and the reason the Trust exists — as the one place AI hasn't reached. NHS England's own flagship rollout proves the pattern is still live: of the named use cases in the 505,000-seat Microsoft 365 Copilot deployment — ward clerks, medical secretaries, HR, finance, procurement — none touch clinical or operational frontline work. Even the boldest national bet so far has landed exactly where the old caution said it was safe to land. ⁴ 

The size of the prize — and whose problem it now is 

This is where the reframe matters: AI is becoming infrastructure, not a project — closer to email or the network than to a piece of clinical software waiting for a business case. 

Infrastructure gets resourced and governed as something the whole organisation depends on. A project waits for a sponsor. 

That reframe is what turns 30% into a number worth taking seriously. Professionally qualified clinical staff — doctors, nurses, midwives, AHPs — number around 750,000 FTE across ⁵ NHS trusts in England. If even a meaningful share of the time currently lost to documentation and admin were recovered, the equivalent capacity works out to roughly 225,000 FTE nationally — over 1,000 additional clinical staff per Trust, at close to zero marginal cost, with no five-to-ten-year training pipeline attached. Set against a 25-year timeline to close the workforce gap by conventional means, that is not a marginal efficiency gain. It is one of the few realistic levers available in that timeframe at all. ⁶ 

Given the scale of that prize, this needs to shift from CIO problem, to jointly owned across the c-suite. 

It is a COO, CMO, and CNO problem first — the people who own the waiting lists, A&E performance, and the workforce gap — and, at that scale, a CEO priority too. Treat this as a roadmap item and it will move at a pace where it never becomes meaningful. Treat it as core to clinical and operational capacity, and it will move at a pace that has transformational impact for patients. 

Why this has to run on two tracks, in step 

AI is not just a technology issue. It is also a people one. 

Its defining feature — natural language, usable by anyone without training in code or data science — is exactly what makes the old sequencing wrong. You cannot govern first and scale later, and you cannot scale first and govern later. Get the balance wrong in either direction, and the programme fails for a different reason each time. 

Move too slowly, and you lose the thing that makes AI different. Heavy governance up front — committees, sign-off chains, pilots that never scale — reads as caution, but it has a cost: staff who tried the tool once, got excited, and then watched nothing happen lose interest, and don't come back for a second wave. 

Move too fast, and you hit a wall anyway. An agent that can't get the data access it needs stalls regardless of appetite. A clinical suggestion that slips past review without the right check causes harm, and one serious incident can freeze an entire programme. Costs that spiral on API or compute use with no one able to explain why erode the case for scaling further, just as surely as a governance committee would. 

Both failure modes point to the same fix: build two tracks in parallel, not in sequence. 

• The operational scaffolding — the practical, unglamorous groundwork: what data the tool can see, GDPR, clinical safety standards, cyber security, and information governance compliance, and the checks that stop a tool drifting across the line into SaMD territory as features are added. 

• The culture change — explicit permission to experiment, training that meets people at their actual skill level, and visible support from clinical and operational leadership, not just IT. 

The two tracks run like a three-legged race, not a relay. 

One leg can be slightly ahead of the other at any given moment — that's normal, even useful, as each pulls the other forward. But the moment one leg sprints far enough ahead to leave the other behind, both go down. 

The decision in front of you 

MHRA has removed the excuse. The regulatory ambiguity that justified confining AI to the back office no longer holds for the clinical and operational admin that eats a third of clinical time. What's left is not a technical question for the CIO — it's a capacity question for the people who own the waiting lists, the workforce gap, and the patient experience. 

Worth being precise about what's settled and what isn't. MHRA's AVT guidance is in force today and is what this argument rests on. The Commission's recommendations are advice to government, not a change in law — a formal response is still pending, with no date attached. Treat the direction of travel as reliable; don't mistake it for the rule. 

The choice in front of every Trust board now is not whether to use AI beyond the back office. It's whether governance and adoption move together, at a pace fast enough to matter, or apart, at a pace that guarantees neither succeeds. 

For a conversation about what that looks like for your Trust, contact Sunny Dosanjh, Hakim Yadi, or Demi Dabor-Alloh.

References 

1. 1. Nuance Communications & Ignetica (independent research consultancy), NHS clinical documentation burden survey, 2022 — commissioned research surveying 966 NHS healthcare professionals across four acute trusts and one mental health and community trust in England (fieldwork April–July 2022). 
2. NHS England (2023). NHS Long Term Workforce Plan.
3. Microsoft and NHS England (2026). NHS England accelerates AI adoption with Microsoft 365 Copilot to improve service delivery, reduce costs and create more time for care.
4. National Commission into the Regulation of AI in Healthcare (2026). Recommendations for a Future Regulatory Framework.
5. NHS England Digital (2026). NHS Workforce Statistics: July 2026.
6. Baringa analysis using NHS England Digital workforce statistics (2026), NHS provider/trust statistics and documentation burden estimates from the NHS Clinical Documentation Burden Survey (2022).

^{6. Baringa Calculations:}

^{1. Total professionally qualified clinical staff (FTE): 755,504 (NHS Digital, July 2026) — used as ~750,000 for simplicity}

^{2. Equivalent capacity nationally: 750,000 × 30% = 225,000 FTE} 

^{3. Number of NHS Trusts in England: approximately 205 (King's Fund, mid-2026)} 

^{4. Per-Trust average: 225,000 ÷ 205 = ~1,098 FTE per Trust, rounded down to "over 1,000" to stay conservative}