Only mad dogs and Englishmen go out in the midday sun.
Or, as Professor Shafi Ahmed put it:
only mad dogs, Englishmen and AI leaders.
Despite last week’s historic heatwave, we managed to gather a whole breakfast table of people willing to brave the London Underground to talk about AI in healthcare. And while the temperature outside was pushing records, the conversation inside stayed cool, calm and grounded.
What followed was an honest discussion about what it actually takes to scale AI in an industry that cannot, under any circumstances, afford to fail. There’s pressure to move quickly, pressure to improve outcomes, pressure to modernise, and at the same time a clear recognition that in healthcare; trust, safety and accountability matter more than ever.
Here are the conversations that stayed with us.
1. The context: healthcare is hard enough, the NHS adds a whole other layer
Before getting into specific barriers, one point sat underneath almost every part of the conversation. Healthcare is already a difficult enough environment for innovation, but the structure of the NHS makes it even harder.
Guests talked about fragmented systems, slow decision-making, uneven digital maturity, complicated procurement routes, and organisations operating under constant day-to-day pressure. There was a strong sense that even when the technology is promising, there is no single path through the system. Different trusts, teams and regions all require different conversations, different approvals and different routes to adoption.
That matters, because knowing your innovation could make a widespread positive impact isn’t enough. You have to know how to navigate a system where responsibility is widely spread, decisions are slowed by committees, and operational reality overpowers any kind of strategic intent.
2. The pilot trap: why promising tools still struggle to scale
There’s a running joke that the NHS has “more pilots than British Airways” (cue the knowing laugh). That doesn’t mean they’re pointless – far from it – but there was a clear sense that healthcare is much better at starting pilots than turning them into anything real.
We mulled over the fact that many AI tools can already show clinical value or strong early results. But what happens next is the hard part: moving through procurement, proving operational fit, finding funding, integrating with existing systems and getting enough organisational buy-in to roll something out properly. The system seems far better set up to run pilots than to turn them into mainstream adoption, and the support structures around it are not necessarily making things any easier. The Health Innovation Networks were raised as one example, with concern that they may not be onboarding enough SMEs into the NHS, relative to the investment behind them. Rather than creating a clear bridge into adoption, the process still feels fragmented, leaving startups to navigate a patchwork of trusts, approvals and local requirements largely on their own. The table’s takeaway: Before you start a pilot, be honest about what comes next. Who would buy it, who would approve it, how would it integrate, and where would the budget come from? If those answers aren’t there, another pilot probably won’t fix that.
3. The most useful use cases may also be the least glamorous
Despite futurist Shafi Ahmed leading the discussion, the examples on the table actually weren’t the most futuristic – they were the most practical. Ambient scribing, imaging and diagnostics, medication support, workflow automation and patient engagement.
Part of the appeal of these tools is that the perceived barrier to entry is lower: there’s still a human in the loop, and the use case is easily tangible. But even here, the room was nuanced. Human oversight helps, but it doesn’t automatically solve workflow, trust, time-saving or liability.
The table’s takeaway: Start with the problems people already feel every day. In healthcare, the strongest early use cases are often the ones that reduce friction, save time or improve consistency in work that already exists.
4. Clinical proof isn’t the same as operational readiness
One of the most useful distinctions raised early on was the gap between clinical evidence and operational evidence. A tool can perform well in a pilot or trial, especially with clean and curated data, and still struggle once it hits the real world.
Real healthcare environments are messier: data comes from different systems, information is missing, workflows vary, and teams are already stretched. Proving the model works is only one step, proving it can survive real operational complexity is something else entirely.
The table’s takeaway: Don’t stop at proving the model works. Test what happens when the data is incomplete, the workflow is messy, the team is busy and the environment is nothing like a controlled trial.
5. Integration is where the hard work really begins
This came up a lot: the AI itself is usually not the hard part. Integration is.
Getting a model to perform is one challenge. Getting it to fit into legacy infrastructure, existing clinical workflows, governance processes, data environments and overstretched teams is another. As one part of the discussion made clear, embedding AI into an existing digital ecosystem can become a much bigger project than building the AI itself.
The table’s takeaway: Design for the systems people already use, not the ideal setup you wish they had. If your product needs major workflow change, heavy IT input or perfect interoperability to succeed, adoption will be much harder.
6. Unclear liability is slowing adoption
A lot of the discussion around liability and accountability was rooted in something very practical: clinicians are understandably cautious about using tools they may be blamed for if something goes wrong.
NHS leaders are often pushed to their limits by immediate, high-stakes crises: workforce shortages, waiting lists, and daily bed management, leaving them virtually no bandwidth to think about digital innovation. That constant firefighting creates a culture where fear thrives. Because leadership is so focused on immediate survival, the prospect of introducing AI (with all its clinical and legal unknown) feels less like an innovation and more like a risk to be avoided. There’s an urgent need for institutions to define the safety profile of their AI tools. If a Trust wants to deploy AI, they’ve got to take the liability off the individual clinician’s shoulders and declare: “We’ll take responsibility for this decision, not you”. Without this top-down assurance, the natural risk-averse nature of healthcare will ensure that the most innovative AI remains on the shelf.
The table’s takeaway: Make it obvious where responsibility sits. If clinicians are expected to use AI but still feel they’ll personally carry the blame when something goes wrong, they will quite reasonably be cautious.
7. Regulation is struggling to keep up
The same disconnect that’s slowing adoption is also showing up in regulation. “AI” is today’s equivalent of a 1950s marketing term: a vague, catch-all label for a huge range of functions. The industry is still struggling with exactly what it’s trying to regulate, because the risk doesn’t lie in the model itself – but in the process, the implementation, and the specific clinical function the tool is performing.
Current regulatory frameworks are struggling to keep up, with recent reports from the MHRA acknowledging that many clinical professionals see existing systems as insufficient for the nature of AI. We’re trying to apply static, point-in-time regulation to technology that’s dynamic, adaptive and highly context-dependent. There was so much emphasis on the need for continuous, real-world monitoring, and for stronger understanding of evaluation and guardrails in practice.
The table’s takeaway: Be specific about the function your tool performs, the setting it will be used in, and the risks it creates in context. That is a much stronger starting point for governance than treating “AI” as a single category.
8. Someone still has to make the call
If there was one practical frustration running through the conversation, it was this: progress stalls when everyone is waiting for somebody else to decide. National policy matters, but so does local leadership, especially in a system where responsibility is spread across trusts, boards, committees and teams.
The reality is that many leaders are already operating at full stretch. But that makes clarity even more important, not less. Trusts and providers that want to move will need to define their own guardrails, set their own expectations, and decide what responsible deployment looks like in practice rather than waiting indefinitely for perfect direction from the centre.
The table’s takeaway: Don’t wait for universal certainty before doing the local work. Someone still needs to define the guardrails, own the decision, and create the conditions for adoption. Build your own “regulatory standard”
Honest conversations matter
What stayed with me from this discussion was how little of it was really about “AI” as technology. The frustration and passion sat around it instead: the challenge of scaling safely, integrating into messy real-world systems, creating enough trust for people to actually use it, and trying to move in a system where responsibility is shared, slowed and deferred.
It wasn’t a debate about hype vs scepticism, but about what it takes to make promising technology work in an environment where the stakes are life and death, and where good intentions aren’t enough.
The opportunity is real, so is the appetite. But if AI is going to scale in healthcare, it’ll depend less on the brilliance of the models, and more on the harder, slower work around them: operational readiness, integration, governance, leadership and trust.
To everyone who joined us: thank you for the honesty, curiosity and candour you brought to the room.
If you’d like to join a future roundtable, drop us a message. We’re keeping the table small, but the conversations big.
