From Access to Trust: What NHS Leaders Told Us About Making AI Work

Transcript from the Breakfast briefing at the House of Lords, hosted by Cohere - 3 November 2026

Whether AI can help the NHS is no longer the question. Who controls it, who trusts it, and how it gets past procurement and governance are.

On 3 November, HETT hosted a breakfast briefing with Cohere and Valarian for a small group of NHS digital, clinical and data leaders: CIOs, CCIOs and CMIOs, GPs, and innovation and population-health leads from trusts and ICBs. The aim was to look past the hype and ask what it really takes to deploy AI safely, at scale and in a way patients and staff will accept.

The answers came as much from the delegates as from the speakers. Here's what we heard.

Phase Two of AI: You Can't Trust What You Can't Control

Jonnie Beddall, VP Strategic Markets at UK AI infrastructure company Valarian, began by zooming out. He was clear that the opportunity is real: using AI "for better clinical outcomes, saving money, certainly, and more efficiency in the system … is not hype."

The barriers are real too: squeezed front-line budgets, a naturally risk-averse system, and the gap between buying enterprise tools at scale and getting real clinical value from them. Above all, he said, there's trust: "the most important word when we're thinking about data." He pointed to recent data breaches and unauthorised access to records as reasons why public confidence can't be taken for granted.

His framing was that the first phase of AI was about getting access to intelligence, and the next is about control: keeping models inside your boundary, and making sure agents can do no more than the person using them is allowed to do.

"If you have access to it, but you can't control it, then you can't trust it."

What it means: data sovereignty, auditability and permissions are now the basic requirements for any AI conversation with the NHS, not optional extras.

Beyond the Chatbot: Judging AI by the Capacity It Frees

Paul Lawrence of Cohere argued that the market has moved on from buying AI for its own sake. The measure now is the time it actually gives back to clinicians and patients: "the way to judge AI is by the care capacity that it creates."

He used AI scribes as a warning. They're popular, but on their own they can add to the load by putting more information in front of staff faster: "sometimes AI scribes just create more work for nurses." The next step is automating the whole process around them. In his example of a shift handoff, notes are drafted, a clinician checks and edits them, and they're written back to the record with sources cited.

One government example showed what that can mean. A workforce-planning report that took a team around ten days, moving between HR, recruitment and vetting systems, now runs in seconds with live data. The team then came back with about 100 more processes to look at, and not all of them were right for AI: "not everything should be done with AI."

Cohere solutions architect Ajay described the building blocks behind this: agents connected to an organisation's own systems, search that works across documents, images and handwritten notes, and ordinary code for calculations, all with a human in the loop. "None of this is a black box," he said. The principle is to "take the models to the data" rather than move sensitive data to the model.

Asked by a delegate how anyone can trust a technology that "tends to want to please", he said answers should be grounded in the organisation's own sources, with citations:

"It is not there to tease you with an answer. It is there to provide you with information grounded in your data sources."

What it means: buyers want to see whole workflows, auditable outputs and honest advice about where AI doesn't fit, not standalone tools.

The Clinician's View: The Gap Is Adoption, Not Capability

Stephen, a consultant neurologist and researcher at Imperial, brought the view from both clinic and lab. He uses AI heavily in his research into how models reason clinically, but almost none of it in clinic.

His point was that capability is no longer the issue. For AI scribes, "there's no longer a question of whether an AI can transcribe and summarise a consultation." The harder questions are which of a dozen credible products is safest and most cost-effective, and whether errors fall unevenly on particular languages, accents or communities.

He was candid about what slows things down. A project his trust ran with technology partners was built quickly, but data-governance approvals took nine to twelve months. Incentives are also out of line:

"The people procuring the systems often don't benefit from them, and so they're only really taking a risk by adopting things."

He added that "not all AI adoption is necessarily good AI adoption", citing imaging triage tools whose false alarms can outweigh their benefit if sensitivity isn't set carefully. What's missing, he argued, is the middle ground: "small-scale pilots" that build evidence and let ideas fail faster, rather than a choice between a finished product and a full trust-wide rollout.

What it means: evidence of safety and equity, along with routes into small, low-risk pilots, may matter more to buyers than another product demo.

What the Room Pushed Back on

Some of the sharpest points of the morning came from delegates.

Look beyond the acute trust. One delegate observed that nearly every example had been hospital-based, though acute care is only around 10% of healthcare activity: "almost all tech companies focus only on the acute … I don't think that's where the future business case is." Another pointed out that a handful of GP systems cover most of England's practices, which gives primary care a simpler route to scale.

Neighbourhood health needs joined-up data. A population-health lead argued that the real intelligence sits across services: "If you can get social care, community, primary care, mental health together, that is a neighbourhood." A community and mental health leader called for quick wins that free up staff time, and for the work to happen "more cohesively as the healthcare system… rather than just in siloed trials." She also reminded the room that patients have to be willing, and that front-line staff are the ones who have that conversation.

Boards need a roadmap. A board-level delegate put the investment problem plainly. Boards are asked to commit public money for five to ten years, but "I haven't seen anybody present to us yet … a common AI strategy for the next ten years." The worry isn't making the wrong decision today, but making a decision today that turns out wrong tomorrow.

Cohere's answer was to avoid lock-in. Its platform is model-agnostic, and the work of getting data and connectors in order keeps its value whichever model comes next.

What it means: suppliers who can talk about primary, community and system-wide care, and about long-term flexibility and predictable costs, will stand out.

Final Thoughts

The mood in the room was neither hype nor scepticism. It was impatience to get it right. Everyone agreed AI can improve care and ease pressure on the NHS. The open questions are about control, evidence, incentives and joining up the whole system.

For suppliers, the lesson is to lead with trust and sovereignty, build evidence through small pilots, think beyond the acute trust, and give boards confidence that today's investment won't be stranded tomorrow.

As Stephen put it, the work now lies in "that gap between the technical capability and what we're using."

Thank you to Cohere and Valarian for partnering on this briefing, and to everyone who joined the conversation.