Hey Insiders - you may have noticed that I skipped an edition. For someone as obsessed about consistency as I am, that was painful.
The truth is that a lot of research goes into writing these and, frankly, every fortnight was a bit ambitious to deliver the level that I think is useful. From here on in I’m going to write these big pieces monthly, and sprinkle in smaller pieces and thoughts through the month (likely to be all the opinions I get part way through writing on Linkedin before the premonition of getting cancelled kicks in).
Please accept this essay as an apology.
Thanks for being here.
James
For most of my time in healthtech, selling AI into the NHS has worked in a fairly predictable fashion. A start-up builds a tool that does a job, it runs a pilot, then tries to turn that into a contract and repeats the process down the road until it runs out of money. The industry calls these point solutions, and the NHS has bought a lot of them.
Ambient voice technologies (AVTs for short) are (were) a good example. So many suppliers piled in that NHS England set up a self-certified supplier registry just to keep track of who met the basic standards. But over the past eight months, a run of announcements suggests that the point solution model might be running out of steam. Platform bundles and joint procurement seem to be in vogue…
February - Epic switch on AI Charting, its own ambient scribe, inside the record system most of the massive US hospitals already pay for. US-based, but you get the point.
March - Manchester University NHS Foundation Trust signs a 3-year enterprise deal with Microsoft covering 6,500 Copilot licences a year and Dragon Copilot (the AVT), plus an internal “agent factory” for staff to build their own automations
April - 4 South West London trusts agreed to put 20,000 clinicians on a single scribing supplier over four years
May - NHS Shared Business Services published a £900m AI framework that runs until 2035
If you’re a founder and your business model right now is selling one tool to one department in one trust at a time, I’m wondering in this piece whether that’s really sustainable. It also raises questions of exactly how you set up to win a joint procurement or get it built around you, but I’ll come to that.
I’ve been hearing versions of the point solution model failing on the podcast for a long time and I’ve tried to test what guests are saying against what the evidence (public contracts and data) actually show. Because if this is a trend, founders and investors need to be aware of what success needs to look like and get ahead of it.
Settle in.
What people are seeing on the ground
A couple of views here from founders at different stages. One founder, who’s building an AI-enabled service for frail older people at home, tried to sell into the hospital discharge pathway, which is the most obvious ROI case in the NHS. Naturally, it didn’t work.
“Everyone acknowledges the huge problem. They want to hear about your solution, but they want someone else to pay for it.”
Sounds familiar. It’s the usual budget boundary story. Hospital says discharged patients are no longer its responsibility, GP says “God, no, don’t send us another alert, because we’re overwhelmed”. Real value, payer didn’t benefit.
Jack, co-founder of Sanctuary Health, an applied AI company that puts engineers inside NHS trusts to build tools the trust then owns, explained how many SaaS technologies are getting commoditised. On scribes specifically, he said pretty bluntly:
“I really want to open source a scribe at some point. I think AVT is a commodity technology. I think it should be owned by the NHS.”
Different vantage points, but the single-purpose tool bought by a single trust is under pressure. Extrapolate and there are a few potential futures here to be aware of…
1. The NHS builds its own stuff
Jack’s thesis. There’s some interesting evidence.
Northumbria Healthcare NHS Foundation Trust actually bought Health Call Solutions in March 2025 so a trust literally now owns software that other trusts use. I’ve also been told that Northumbria runs a large in-house engineering function building software it can resell to other trusts. I haven’t verified the size of the team, so don’t hold me to the specifics, but the direction’s pretty clear and very public.
However… there’s a healthy case for betting against this future.
In Bessemer’s 2026 Healthcare AI ROI Scorecard (survey of US health orgs), 61% said half or fewer of their internally built AI tools are actually maintained. Maintenance of AI tools is easily forgotten and anyone with more than a couple of scheduled Claude tasks and/or a Zapier account will testify to an automation never being fully solved and forgotten. As a result, effort’s apparently moved away from internal builds to specialist vendors. Now, let’s do some journalism here and appreciate that Bessemer is a VC firm, they have a healthcare SaaS portfolio to hype up, and US hospitals aren’t NHS trusts. Still, it’s data that doesn’t flatter the ‘build-it-yourself’ case.
Okay, but would the NHS open source a scribe? NHS England made all of its source code repositories private by default from 11 May 2026, saying that AI models analysing public code at scale = potential apocalypse, so it doesn’t feel like it.
NHS England also told trusts to stop using non-compliant ambient voice products, and made the trust, the practice or the individual clinician liable. So if you build your own stuff, you own that risk. I imagine that, alone, is a strong enough reason to buy rather than build.
So building in-house is happening in pockets. But it doesn’t look like the main force. Not in the short term, anyway.
2. Big platforms swallow entire categories of AI solutions
Here are 3 examples at different stages…
Clinical notes. When Epic switched on AI Charting, the press called it the moment all standalone scribes got “Sherlocked” (a platform shipping a feature that kills a paid add-on). Why buy something standalone if it’s part of something you already pay for and integrate into your organisation? Good argument.
The UK version of this is Microsoft. NHS England’s £10bn, three-year technology package includes Copilot for more than half a million staff and a national push to scale AVT, starting with outpatients. If you look at Manchester as precedent here, once AVT arrives as part of an enterprise agreement or a national programme, all standalone suppliers then have to beat something that looks free and without any irritating information governance, regulatory or other red tape issues. Even the South West London deal, which went to Lyrebird, made a point of the integration with Oracle.
Hospital operations. This is where the Federated Data Platform comes in. If you missed it, I went through the FDP in detail in WTF is the FDP?, so I won’t go over the whole saga again. The TLDR is that it’s one piece of software running separately inside each trust, with NHS-owned products on top for things like theatre planning, discharge and waiting list validation. In March, adoption was 86% of acute trusts, with plans to add AI-assisted theatre scheduling, AI-supported discharge summaries and a “Demand Centre” for triaging referrals (which had already triaged 26,544 referrals in North West London).
Now, theatre scheduling, discharge and referral triage are all categories start-ups have historically sold trust by trust. But here’s the twist…
