Finn Brooks: Clara, long week — I had to actually call my GP surgery this morning and I sat in a phone queue for eighteen minutes before I gave up.
Clara Bennett: Eighteen minutes to give up, not eighteen minutes to get through.
Finn Brooks: To give up! And look — that is, genuinely, the exact broken system that today's episode is built around. Because NHS England just announced, on July 4th 2026, the NHS's seventy-eighth birthday, that they're deploying AI triage directly inside the NHS App. The whole pitch is: instead of that phone lottery, the app asks you adaptive questions — your symptoms, how long, any history — and tells you whether you need a GP, a pharmacist, A&E, or just rest. No queue. No guessing.
Clara Bennett: And that sits inside a ten-billion-pound government commitment to overhaul NHS technology over three years.
Finn Brooks: Which sounds enormous — until you look at the evidence underpinning the actual rollout. One trial. One place. Wealden Ridge Medical Partnership, a GP practice in Sussex, saw a twenty-nine percent drop in patients queuing on the phone. That's it. And now we're talking about two hundred thousand patients within twelve months, every NHS App user by April 2028.
Clara Bennett: Hold on — I want to make sure we're being fair to what that number actually shows. A twenty-nine percent reduction in phone queues tells us the phone system was a catastrophic bottleneck. It doesn't — and this is the key distinction — it doesn't tell us whether the routing decisions the algorithm made were clinically sound.
Finn Brooks: No, that's — yeah, that's exactly the gap I can't get past. So how did one Sussex surgery become the whole foundation for a national policy?
Clara Bennett: That's the question. And I think the answer is uncomfortable.
Finn Brooks: Okay but — the uncomfortable part is *who* ran the trial. Dr Ragu Rajan, a clinician at Wealden Ridge Medical Partnership, was involved in designing and running it. So you've got one practice, one clinician-investigator, rural Sussex — and that's the data point the whole national commitment rests on.
Clara Bennett: One data point. And to be clear, that's not a criticism of Dr Rajan's work — the question is what a single-site trial can actually prove. Wealden Ridge is rural Sussex. Which means, in practice, you're looking at a patient population that is almost certainly more digitally confident, more likely to have a smartphone, probably more affluent, than an urban NHS trust in Birmingham or Tower Hamlets.
Finn Brooks: So we have one data point and we're calling it a floor.
Clara Bennett: That's exactly it. And NHS England South East Region did publish a report in October 2025 documenting AI deployment across general practice — but documenting *spread* is not the same as controlled validation. It tells you the tools are being used. It doesn't tell you they're routing people correctly.
Finn Brooks: Wait — so there's a regional report, it's October 2025, and it's basically just — here are the places using AI now?
Clara Bennett: Spread documentation, yes. Now, that's useful evidence that adoption is happening — but it's not a controlled trial measuring whether the triage decisions themselves prevented harm or caused it. Those are completely different claims.
Finn Brooks: And that's — I mean, the twenty-nine percent reduction is real. Wealden Ridge is a real place. But what it actually measured was whether fewer people were stuck on hold. Not whether the algorithm looked at someone's chest tightness and said the right thing.
Clara Bennett: In practice, what you've proven is that the old system was a terrible bottleneck — which, yes, absolutely it was. But demand management as a strategy — redirecting patients away from GPs toward pharmacies or self-care — only works if the routing is clinically accurate. A phone queue metric tells you nothing about that accuracy.
Finn Brooks: So the baseline is broken enough that almost anything looks like progress — and that might actually be the most alarming part of this whole story.
Clara Bennett: Because the alarming part isn't just the low bar. It's that nobody has publicly stated who is responsible when the algorithm is wrong.
Finn Brooks: Wait — like, there's no framework at all?
Clara Bennett: No publicly stated clinical governance framework. None. So imagine — someone wakes up, chest tightness, opens the NHS App, answers five adaptive questions, gets routed to self-care. Two days later she's in A&E with a cardiac event. Who bears that? NHS England? The algorithm's developer? The GP whose patient data trained the model?
Finn Brooks: And the answer right now is — genuinely nobody knows. That accountability gap is just sitting there, open.
Clara Bennett: Right — and the tool can fail in both directions. Under-triage a serious condition, you get that cardiac scenario. Over-triage a minor one, you flood A&E with people who didn't need to be there. Neither error threshold is publicly defined.
Finn Brooks: Okay but — wait, this is the part that actually, I mean — the comparison problem makes it worse, right? Because the 8am phone lottery is itself considered unsafe. So we're not measuring the AI against a safe system. We're measuring it against a broken one. And James Murray stood up on the NHS's seventy-eighth anniversary and endorsed this — that's symbolic framing, not a safety threshold.
Clara Bennett: That's — yes, and the health bodies that broadly welcomed the rollout simultaneously called for the NHS to prioritise patient safety, confidentiality, and inclusion. But none of those calls have translated into binding governance. It's cautionary language without teeth.
Finn Brooks: So the caution is real but toothless.
Clara Bennett: And the inclusion piece — that's actually the thread we need to pull next, because the NHS App as the gateway here may be excluding the exact patients who need triage most.
Finn Brooks: The gateway problem — because that's what it is, right? The NHS App is the *only* door into this system. There's no phone version, no walk-in version —
Clara Bennett: That's the hardest fact in this whole conversation, actually. The NHS App is the sole patient-facing platform for this feature. And the Wealden Ridge trial — a rural Sussex practice where patients, by definition, found and used the app — gives us zero data on how an elderly patient in Bradford, or someone who arrived recently and speaks Urdu as their first language, would navigate this.
Finn Brooks: Zero. That number is actually — wait, that's not a gap in the evidence, that's an absence of evidence entirely.
Clara Bennett: Completely unquantified. And health bodies specifically flagged inclusion — alongside safety and confidentiality — as a concern. That's not incidental. That's the sector signalling the current design does not adequately address this. Now, the counterintuitive part: NHS England is simultaneously using the AI notetaking result — clinicians spending up to twenty-five percent more time with patients — to justify accelerating the whole programme, including triage. The logic being, if AI is already delivering, deploy more AI faster. But the patients gaining that extra clinical time are the ones who got *through* the door.
Finn Brooks: Oh that's — hang on. The people who benefit from the better appointment are already app-users. So the tool widens the gap it was meant to close.
Clara Bennett: That's the mechanism, yes. And separately — the AIR-SP platform, nearly six million pounds in NHS AI screening infrastructure — that's running across multiple care pathways at the same time. So parallel systems, parallel investment, and the inclusion risk is unquantified across all of them.
Finn Brooks: And then there's the Cobalt Park thing — OpenAI apparently never visited it. This is the North Tyneside site that was designated as an AI growth zone, central to the thirty-billion Stargate UK pledge. Never visited.
Clara Bennett: Which is — mm, that detail matters because it's the same pattern. Headline commitment, announcement energy, and then the on-the-ground delivery just... doesn't materialise in the places it was promised.
Finn Brooks: So the NHS App only opens for people who already know how to use it, the data proving it works comes entirely from those people, and we're calling that a foundation for two hundred thousand patients by next July.
Clara Bennett: And that's the part I actually can't settle. April 2028 isn't a checkpoint — it's a commitment. There's no stated off-ramp. If safety signals emerge at scale, at two hundred thousand patients, at a million — the mechanism for pulling back just hasn't been described publicly.
Finn Brooks: And I don't know — I mean, I genuinely don't know — whether NHS England has visibility into which way this lands, or whether the political cost of not acting has just... overtaken that question entirely.
Clara Bennett: That's the honest version of it, isn't it. The broken baseline might be the whole argument. When the 8am phone lottery is already causing harm, the calculation becomes: is imperfect AI less dangerous than the status quo. But that calculation hasn't been made public. We don't know if anyone has actually run it.
Finn Brooks: Yeah. That's where I keep getting stuck.
Clara Bennett: It's a genuinely open question. Thanks for working through it.
