AVT in the NHS: a rare chance to get it right

In April I joined a King's Fund roundtable on scaling ambient voice technology (AVT) in the NHS, alongside NHS leaders, patient representatives and other suppliers. The King's Fund has since published its long read from that discussion, and I'll be picking up where it leaves off at their follow-up conference, Ambient Voice Technology: what's next for health and care?, on 12 November.

One concern ran through the whole room: how do we stop AVT going the way of so much NHS technology before it? Electronic patient records and dictation software arrived with real promise and real money behind them, but the change needed to make them stick got skipped, so the benefit landed in patches rather than across the system. The King's Fund's warning is blunt: AVT will repeat that pattern if the NHS buys the technology without funding the transformation around it. 

I came away more optimistic than that sounds, because the enthusiasm for AVT among frontline staff is genuinely different from anything I've seen with NHS tech before. But enthusiasm isn't a plan. Here's where I think the real work sits.

Focus on the change, not just the tool

The single biggest lesson that the public sector has learned over the last decade is that the technology is rarely the hard part. Trusts buy a tool and the rollout stalls, because there's no dedicated capacity to manage the change, train clinicians and rebuild workflows around it. Tools implemented without clinician involvement see lower adoption and less impact, whatever the product itself can do.

That's an opportunity for suppliers to approach this differently, us included. NHS organisations should be able to expect robust implementation and lifecycle support as part of the product, not a licence that's sold and then left to run itself. In practice that means co-designing the rollout with clinical leadership and local clinical teams, embedding champions on the ground, adapting to how each team actually works, and staying heavily involved well past go-live. Skip that, and you've bought technology nobody uses.

The real prize is Phase 2

The report sets out a simple way to think about AVT maturity. Phase 0 is transcription, turning speech into text. Phase 1 adds summarisation and basic integration into the record. Phase 2 moves into advisory coding and suggested actions. Phase 3 is full clinical decision support.

Most tools on the market today sit around Phase 1. The King's Fund's view, backed by the clinicians it spoke to, is that the real time savings don't show up until Phase 2 - when a conversation doesn't just get summarised but turns into the tasks, codes and next steps a clinician actually needs. That's exactly the direction our Notes work is already heading, pulling together whole assessments from multiple sources rather than drafting up a single consultation. It's also a useful marker for where the sector as a whole needs to go, and quickly.

A high bar on evidence - shared, not endlessly repeated

The report calls out a real gap: there's no standardised way to compare AVT suppliers on the things that matter, like error rates, reliability and uptime. Without that, buying decisions risk coming down to price alone - something we're already starting to see. No single supplier, us included, should get to mark its own homework, so the fix has to be collective: shared metrics, shared reporting, clear benchmarking. I'm firmly in favour of holding suppliers to a high bar.

The harder part is doing this proportionately. There is already a lot of evaluation happening around this technology, and the risk is that we spend all our time evaluating and none deploying. When the tenth trust runs the same foundational accuracy and safety pilot that nine others have already run in the same setting, in isolation, that’s duplication that slows everyone down. The answer isn’t to test less rigorously - it’s to test rigorously and then share what’s been demonstrated, so each organisation can focus its own effort on what's genuinely local: workflow fit, team integration, how it integrates with the systems already in place.

It's also true that these models are changing fast, which is a fair argument for not treating any single evaluation as a permanent tick in a box. But that points towards continuous, transparent monitoring that everyone can see, not towards every trust re-running the same from scratch. Rigorous, proportionate, and shared: that's the balance I think we need to strive for.

Patients should shape this, not just be told about it

Patients came up throughout the discussion, and rightly. People want to know, in plain language, when AVT is being used, and to have the choice to opt out for sensitive conversations.

I'd go a step further. Patients and the people accessing services should help shape these tools, not simply be informed about them after the fact. That’s how we build at Beam. We started out as a services organisation delivering on the frontline, and we’ve carried that through: our technology is built by frontline workers, for frontline workers. Just last year, on the social care side we ran research with residents and people with lived experience, in partnership with Nesta. We hold ourselves to the same standard in healthcare. Done well, that isn't the slow route - it's how you get it right the first time, rather than reworking later.

AVT shouldn’t stop at the NHS boundary

There's a wider shift worth naming. NHS England is being absorbed into the Department of Health and Social Care, and the push towards neighbourhood-level, joined-up teams means that health and social care will have to work more closely together than ever before. AVT is already being implemented across both - so any structured data layer or shared platform that AVT makes possible needs to serve both sides of that boundary from the start. We've seen a version of this in our own work with NHS Hampshire and Isle of Wight, where joining up a complex CHC assessment pathway was where the difference was made. Right now most of the AVT conversation, this report included, stops at the NHS boundary. That feels like a missed opportunity given where the whole system is heading.

Sovereignty is easy to say and hard to procure

UK sovereignty is a live conversation, and rightly so given how much sensitive patient data is involved. My view is that it’s less about where a supplier's head office sits, but rather what happens to the data: where it's stored, where it's processed, and whether it trains someone else's model. For our UK customers, data is stored in the UK, with some processing in the EEA, and is never used to train models - details we publish in full on our Trust Center. But that's fast becoming basic hygiene.

The more interesting question is groundwork. A handful of companies, many of them British, have spent years doing the unglamorous work of proving what AVT can actually do in real frontline conditions - piloting, learning and iterating. The risk is that the moment a category is de-risked, it gets treated as a commodity - bought on price, as if every tool were interchangeable - when the years of groundwork are precisely what make them not. Backing those companies is a policy choice as much as a market one.

AVT is the starting point, not the destination

If there's one point I want to land, it's this: AVT is where this begins, not where it ends. Transcription and summarisation are quickly becoming table stakes. From our experience deploying to over 100,000 users across health and social care, the pattern is consistent - partners adopt AVT, see the value, and almost immediately ask, "what else can you do?"

That question is the interesting one, and for us at Beam it points in two directions. The first is whole workflows rather than single tasks: not better notes at one appointment, but AI that supports the full patient journey, from referral through assessment, MDT discussion and treatment to discharge and follow-up. The second is new modalities beyond documentation altogether. Real-time interpretation for patients who don't share a clinician's first language, which is a genuine equity issue, and something we’re already building towards with Beam Interpret. And voice agents - our own is Beam Talk - can take the pressure off the front door, screening common queries, gathering information, and keeping in touch with people on waiting lists. 

Getting there isn't something suppliers can do alone, and it isn't something the NHS can specify from the outside. It comes from partnership - working together from the earliest stages, with the NHS shaping the product using the one asset no supplier can replicate in a lab: real clinical workflows, real edge cases, the challenging reality of delivering care at 3am on a Saturday night in a community mental health team. So the ask runs both ways. Buy the change alongside the technology, share evidence instead of duplicating pilots, and partner with suppliers who co-build with clinical teams and patients - and wrap implementation support around the product - rather than selling a licence and walking away. Get that right, and it's less a purchase to be made than a partnership to be built.

I'll be exploring all of this in more depth at the King's Fund's follow-up conference, Ambient Voice Technology: what's next for health and care?, on 12 November - I'd love to see you there.

About the author: Matilda Crowfoot is the Strategic Lead for Healthcare at Beam, the social impact technology company behind AI tools now used by over 100,000 frontline staff across health and social care. She heads up Beam's partnerships with NHS and private healthcare organisations, helping them implement these tools in frontline care settings and realise the benefits in practice, from adoption through to long-term impact. Prior to Beam, Matilda worked in global health and wellbeing consulting, designing and implementing large-scale programmes across the public, private and social sectors internationally. Her experience spans behaviour change, health innovation and systems-level transformation.