There’s a quiet revolution happening in NHS waiting rooms, or rather, happening before you even get to one. AI-powered symptom checkers and triage tools are being piloted across dozens of NHS trusts in England right now, and depending on who you ask, they’re either the future of healthcare or a genuinely worrying experiment conducted on millions of patients who didn’t really sign up for it. I’ve been following this story for a while, and my honest view is: it’s both.
What NHS AI triage actually looks like in practice
The most visible example is NHS England’s rollout of AI-assisted triage through platforms like Babylon Health’s GP at Hand (now much reduced in scope), Kry, and the symptom checker integrated into the NHS App itself. The NHS 111 service has also been testing AI tools that help call handlers prioritise patients before a clinical assessment takes place.
In practice, a patient opens an app or visits a portal, types or speaks their symptoms, and the AI runs them through a decision tree that’s been trained on clinical datasets. The output might be: “attend A&E within the hour”, “book a GP appointment within 48 hours”, or “self-care at home.” These aren’t random guesses, the better systems are trained on genuinely enormous pools of anonymised patient data. But the margin for error, when that error involves a human being, is a very different kind of problem than the margin for error in, say, a pothole-detection algorithm.
The clinical evidence: promising, but patchy
The data supporting these tools isn’t bad. A 2023 study published in The Lancet Digital Health found that the Ada Health symptom checker correctly identified the right condition in the top three results for roughly 71% of test cases, which sounds impressive until you consider that a GP, with full patient history, the ability to ask follow-up questions, and basic physical examination, gets it right considerably more often. The MHRA, which regulates medical devices including software, has been updating its guidance on AI tools under the Software and AI as a Medical Device framework, acknowledging that the evidence base is still being built in real time.
The Nuffield Trust, a well-regarded health policy research charity, has pointed out that the clinical trials underpinning most of these tools were not conducted on the full diversity of NHS patients, older people, those with multiple long-term conditions, or patients for whom English is not a first language. That’s a significant gap when you’re deploying tools nationally.
What GPs are actually saying
I’ve read through a fair amount of GP feedback on these pilots, and the honest summary is: GPs are not uniformly opposed, but they are concerned about specific failure modes. The British Medical Association has raised worries about liability, if an AI triage tool tells a patient to stay home and that patient deteriorates, who is responsible? The GP who was bypassed? The trust that deployed the software? The company that built the model?
There’s also a subtler issue around what clinicians call “safety netting”, the practice of telling a patient what warning signs to watch for and when to come back. AI systems are getting better at this, but the conversational texture of a GP saying “if you develop a fever above 38°C or the pain spreads, go straight to A&E tonight” is harder to replicate than it sounds. A static screen of bullet points doesn’t carry the same weight, and there’s evidence that patients process it differently.
On the flip side, several GPs I’ve seen quoted in Pulse magazine have acknowledged that the NHS is drowning. With roughly 8.1 million people on the elective waiting list as of early 2026 according to NHS England’s own figures, any tool that safely filters out self-limiting conditions frees up appointment slots for people who genuinely need them. That’s not a trivial benefit.
The patient safety concerns regulators are taking seriously
The Care Quality Commission has flagged that some AI triage products being used in NHS settings are operating in a regulatory grey area, technically classed as decision-support tools rather than medical devices, which means lighter-touch oversight. The MHRA is working to close that gap, but regulatory frameworks do tend to move slower than software deployment timelines.
Algorithmic bias is a live concern too. Research from University College London found that some symptom-checker models perform less accurately for women describing cardiac symptoms, partially because the training data historically over-represents how men present with heart conditions. The classic “elephant in the room” problem: your model is only as fair as the data it learned from.
There’s also a real tension between the drive to reduce A&E pressure and the risk of under-triaging. An AI that’s calibrated to be conservative (sending more people to hospital to avoid missing serious cases) adds to emergency department queues. One calibrated to be liberal (keeping people at home) risks missing things. Getting that dial right is genuinely hard, and it may need to be set differently for different population groups.
Is this actually different from what we already do?
One argument I find compelling, even if it makes GPs uncomfortable, is that the status quo isn’t safe either. NHS 111 has always involved non-clinicians asking scripted questions and following decision trees, the AI version is arguably just a faster, more consistent version of that process. And the data from NHS 111 online (the web-based symptom checker that’s been running for years) shows it handles around two million assessments a month with a broadly acceptable safety profile.
That said, the ambition of current pilots goes further. Some trusts are testing AI tools that analyse GP notes to flag patients at risk of deterioration before they even contact the NHS. That’s a very different proposition from a symptom checker, it’s predictive medicine, and the ethical questions around it are considerably thornier. It connects, in my mind, to broader conversations we’re having about AI making decisions that affect people’s lives, from fitness and wellness apps all the way up to tools that inform clinical pathways.
The NHS AI triage app landscape in the UK is moving fast. The technology is genuinely impressive. The clinical validation is improving. But the governance frameworks, the liability structures, and the work needed to make these tools equitable across the full diversity of NHS patients still have a long way to go. I’d keep a close eye on how the MHRA’s updated AI medical device regulations land later this year, that’ll tell us a lot about whether the NHS’s bet on this technology is a calculated risk or a leap of faith.