The risk in NHS ambient voice technology contracts is year four

In March 2026 Hertfordshire Community NHS Trust signed a three-year contract with Accurx for ambient voice technology across more than 1,000 clinicians. Ambient voice technology is the AI notetaker sold as a scribe: it listens to the consultation and drafts the clinical note. The same product now runs in more than 40 trusts, and since July it is national policy, with NHS England backing a rollout that starts in outpatients. Hertfordshire is the example here because its announcement is unusually clear about what was bought: headcount, appointment volume, term. I am not suggesting it did anything wrong. I am pointing out what the announcement, like every other one I could find, leaves out. It says nothing about what the price does when the three years end.

That silence is the subject of this piece. Inside the term, a trust is fine; whatever was agreed in March is what it pays. The question is year four.

One hop behind the contract is a price the trust never sees

The trust buys from Accurx. Accurx's product is built by Tandem Health. Tandem runs it on Microsoft's cloud. We know this because an NHS commissioner said so in a Freedom of Information response: the solution "uses the open source model Whisper and Azure Open AI". Whisper turns speech into text. Azure OpenAI is the service that turns the text into a clinical note, and Microsoft's own pricing page describes how it charges: pay as you go, per token, a token being a short chunk of text. There is a reserved tier that fixes a rate, but only for a month or a year at a time, and a reservation renews at whatever the price is then. Nobody has published which tier Tandem is on.

What we do know is that the price moves, and that the people it moves for are not in the NHS. From 1 September Microsoft raised its regional prices: a reservation in Sweden, where the FOI response says Tandem is based, now sits 30% above Microsoft's global rate, and one in the UK sits 40% above. In August OpenAI cut a flagship model by 20% on input and 33% on output, for at least three months, with nothing said about what follows. The same month Anthropic cancelled a scheduled 50% rise three weeks before it was due. Prices in this market fall as often as they rise. That is not the point. The point is that a trust on a three-year fixed price has no idea which way its supplier's costs have gone by the time it comes back to the table, and no clause anyone has published ties the two together in either direction.

Compare that with how the NHS treats money it does control. NHS England's capital guidance gives trusts a firm envelope to 2029/30, with multi-year operational capital going direct to providers "for the first time", and says in so many words that this is "designed to provide the confidence needed for long-term investment decisions". Four years of certainty on the buildings. On the software inside them, a fixed term that ends in a market nobody in the NHS sets.

The half you can price is the cheap half

I tried to put a number on it, and the exercise is instructive mainly for where it stops. Transcription has a published price: Microsoft lists Whisper at $0.4356 an hour of audio in the Sweden Central region. Run Hertfordshire's own 250,000 appointments a year through it, at the closest published UK consultation lengths, and transcription costs the trust something like $16,000 a year. For a three-year, thousand-clinician deal, that is nothing. Scale it to the one part of the NHS that publishes how long its appointments last, general practice at about 386 million appointments a year, and the transcription bill for the whole of English primary care is somewhere around $38 million. An illustration, not a forecast: nobody transcribes every appointment, and a third of them have no recorded duration at all.

But notice what that sum covers. Only the layer that writes the transcript. The layer that writes the clinical note, the proprietary model behind Azure OpenAI, has no published rate for whichever model Tandem runs, and a structured note plausibly costs more to generate than the audio costs to transcribe. The half of this bill that is easy to price is not the half that matters. And England records roughly 663 million clinical contacts a year while publishing durations for one setting out of five, so the sum cannot even be attempted for outpatients, mental health or community care. The only national modelling anyone has done covers emergency department doctors and rests on its own stated assumptions.

It is the same split I wrote about in buy the commodity, build the moat. Whisper is open source. It costs money to run, and the Azure rate above is what Microsoft charges to run it for you, but a trust could in principle run it on its own hardware, at a fixed cost, and keep any saving as models get cheaper. The note-writing model is the other kind: proprietary, metered, priced by someone else. The exposure in this pipeline is not spread evenly. It sits entirely on the half that cannot be owned, and that half was chosen, across the whole NHS, without anyone arguing for it. Cloud was what you could buy.

None of this is an argument against ambient voice technology. Clinicians want it and the early evidence says it gives time back. It is an argument for reading the renewal clause before signing one. Inside the term the vendor carries whatever the backend does. At renewal the vendor decides how much of it to carry forward, and that decision is the one nobody has published a rule for. Hertfordshire will be back at the table in 2029, and so will forty-odd other trusts, in a market that repriced itself three times in one summer.

So three questions, for anyone signing one of these this year. Which tier does the deployment run on, metered or reserved? Does any clause tie the trust's price to what the backend costs, in either direction? And what, in writing, happens when the term ends? I could not find a single published board paper, procurement notice or contract that answers any of them. Nobody is being overcharged, as far as anyone can tell, and "as far as anyone can tell" is the problem.

The NHS knows how to buy itself four years of certainty. It has just done it, for the buildings. It should ask for the same certainty on the software inside them before it signs.