Almost 800,000 people were waiting for NHS treatment in Wales as of May 2025, according to the British Medical Association - and it’s a multi-year, recurring problem, not a blip. The Care Quality Commission put it plainly in October 2025: demand for care is rising across a health and social care system that’s already under severe pressure. NHS England regularly cites missed appointments as a £1 billion-plus annual cost. And workforce and capacity constraints mean, in the House of Commons Library’s words, that demand is simply outstripping supply.
Those numbers framed a recent webinar on scaling patient access with AI agents, hosted by Robotics AI and Druid AI, featuring two clinicians who deal with that gap every day: Dr Mike Brady, Assistant Clinical Director for Clinical Care at the Welsh Ambulance Service, and Dr Tim Caroe, Medical Director of Primary Care and acting CCIO at NHS England South East. What they described wasn’t a pitch for AI as a fix-all - it was a working account of where AI agents actually help in urgent care, where they don’t, and what it takes to deploy one safely inside an NHS trust.
It’s not an access problem — it’s an information problem
Asked where he sees the biggest breakdown in patient access, Brady pushed back on the framing entirely: “I don’t necessarily think it’s an access issue, I think it’s an information before access issue.” Patients - even ones like him and his paramedic wife, who know these systems inside out - struggle to work out which service to use, for whom, and when. “We’ve kind of... put the onus on people to have to make what is quite a complex decision when they’re having their worst day, their worst moment,” he said. Years of “don’t go there, don’t call here” messaging haven’t solved that; they’ve just shifted the burden onto patients at the worst possible time.
His conclusion: give people more information before they decide how to access care, in the format that works for them, and a good share of the apparent “demand problem” turns out to be a routing problem - peaks that could be flattened by helping people reach the right point of contact the first time.
Dr Tim Caroe extended the point to NHS England’s own roadmap: the NHS App is being built toward a single digital front door that can ask “I don’t feel well - in what way?”, gather information conversationally, and signpost across channels - the app itself, a smart speaker, WhatsApp, a desktop browser. But both were emphatic that this isn’t about replacing the phone line. As Caroe put it, “not everyone is digitally savvy, and we must make sure that we don’t leave people behind... one of the great things about digital is it frees up space... so that you are available to pick up the phone and have that conversation with someone who can’t use the digital tools.”
Co-design, not top-down design
Asked how he’d redesign the front door to care if he could start from scratch, Brady’s answer was disarmingly honest: “I wouldn’t redesign it, not least on my own.” His concern was that a solution designed by clinicians for an assumed “average” patient would fail the people with additional or specialist needs. “There’s no point in me designing something that I think is going to work for the largest part of the population” if it excludes people it should serve. Person-centred, co-designed access - built with the people who’ll actually use it - was, in his words, “key.”
The case: AlBot on NHS 111 Wales
While a longer-term programme to overhaul its website continues, the Welsh Ambulance Service deployed an agentic AI assistant - nicknamed AlBot by staff, via an internal competition - on the NHS 111 Wales website, built with Robotics AI and Druid AI. Brady was candid that the value wasn’t just the good headlines; he walked through the real lessons.
Hard guardrails on source material. AlBot only draws on the NHS 111 Wales website - nothing else. “We don’t want agents going off and finding information that’s unregulated, that’s unsupported, and that’s written in very complicated ways,” Brady said. Using RAG, it matches plain-language queries - “ouchy ear” works as well as “otitis media” - to the right page and returns the Trust’s own approved guidance, in multiple languages, without interpreting it.
It doubles as service-design intelligence. Beyond helping patients, AlBot’s query data tells WAST what people are actually searching for - informing which website pages and services need attention first.
The hardest lesson: sensitive topics. Internal testers - people modelling the range of users the site serves - deliberately tried to break it, and the sharpest gap they found was how it handled sensitive disclosures, like someone indicating they might self-harm. Where the knowledge base had no direct answer, the bot’s fallback (“I haven’t got enough information about that right now, sorry”) was, in Brady’s words, “not good enough... not compassionate enough” and failed to escalate or signpost. The fix mattered as much as the finding: instead of treating this only as a “deflect demand” tool, the team built it to actively encourage contact in these moments - directing people to call, every time, for exactly the situations where a scripted answer would be the wrong answer.
Tone, transparency and regulatory boundaries. The team worked hard on tone and language so it was unmistakably clear AlBot “is not a human” and isn’t interpreting information, only retrieving it - a distinction that matters both for patient trust and for staying on the right side of medical device regulation. Getting that right, and making sure staff across the organisation understood it before launch, was - in Brady’s words - “quite a challenge, but well worth doing,” because you wouldn’t deploy this live without knowing it gives safe, branded, trusted information.
Governance has to be a first-class citizen. Data privacy impact assessments, clear data processor/controller lines, explicit prompts asking users not to share personal information, and - notably - quality and equality impact assessments were all built in from the start, with information governance and clinical safety colleagues involved from day one. Brady’s framing: “we can probably do the most for the most” - improving access for the majority frees up capacity and space for those who need a different, more tailored form of support.
Start small, low-complexity, high-volume
Don Slope of Druid AI summarised the deployment philosophy simply: “we’re not trying to boil the ocean... let’s just focus on something that’s low complexity, high volume, and let’s start there.” Brady confirmed that’s exactly the path WAST took - and that public confidence has grown enough since launch that the Trust is now extending the approach into other service areas.
Throughout the conversation, both clinicians returned to the same point: AI should create capacity without compromising dignity. By automating routine interactions, clinicians can focus on patients who need nuanced, human judgment - particularly those managing multiple long-term conditions, where standardized guidelines often fall short.
What would unblock adoption across the NHS more broadly
Asked what NHS England could do to remove barriers to adoption, Caroe pointed to two structural levers: pushing electronic health record suppliers to open up standardised APIs so agentic workflows are possible by design, not as a workaround; and national-level assurance frameworks, so individual GP practices and trusts aren’t left to independently evaluate the safety of complex AI systems on their own. He also noted a cultural shift already underway, citing NHS England chief executive Jim Mackey’s message to the service: “let’s just get on with stuff” - don’t let two years of caution get in the way of benefits the evidence already supports.
The takeaway
Scaling patient access in urgent care, on this evidence, isn’t primarily a technology procurement question. It’s a design and governance discipline:
- Treat information as the first access barrier
- Co-design with the people who’ll actually use the system
- Start with a narrow and well-understood use case
- Build escalation paths for the moments that matter before you launch
- Use what you learn to extend cautiously rather than all at once
That’s the model the Welsh Ambulance Service and NHS England described — and it’s a more transferable blueprint for any urgent care organisation than a features list ever could be.
“It’s like having a buddy your entire shift.”
