Ben Turner | Medical AI Product Lead, Heidi, London, England
Citation: EMJ Innov. 2026; https://doi.org/10.33590/emjinnov/8A40BAM2
Your work spans clinical practice, AI product development, and the implementation of AI across healthcare systems. How has working across those different settings shaped the way you think about designing AI that genuinely improves clinical care?
My background is in vascular surgery, and I founded a company called AutoMedica, London, UK, which built point-of-care AI medical knowledge tools before Heidi acquired it. So, I came into this job from both directions at once. When practising clinically, I felt the administrative load AI tools are supposed to lift, and now I’ve spent years on the other side building the products meant to lift that same burden. A typical week has me sitting in on clinical sessions, then in a room with our engineering team discussing how Heidi’s AI Care Partner should behave in a consultation, developing the safety harness to evaluate a new AI product before moving on to a call with a health system explaining how we ensure clinical safety and efficacy during development and beyond into real-world clinical use. That combination is what lets me bring both a clinical and a product lens to the role at Heidi. Decisions made purely in a product meeting, away from a clinician or end-user, tend to be confidently wrong.
When you introduce AI into the consultation process, even in a supporting role like administration, it inevitably changes the dynamic of the clinical encounter. What shifts in behaviour or communication have you observed between clinician and patient when these tools are used in practice?
The thing I hear about most is eye contact. A clinician typing notes during a consultation is, for a good chunk of it, looking at a screen rather than the person in front of them. Take that away and something changes in the room almost immediately. General practitioners (GP) have told me they catch themselves listening differently once they’re not simultaneously composing a sentence for the record in their head. Patients notice it too, usually before they could tell you why. Some clinicians have also started narrating what’s happening, mentioning that a summary is being generated so they can look over it together at the end, and that small habit does more for trust than any amount of small print. For me personally, it leads to explaining the full plan and next steps to the patient and ensuring I don’t miss anything in my documentation, so it also leads to more open and thorough communication.
There is increasing attention on patient awareness of AI being used during consultations, particularly in documentation and summarisation. From your experience, what do patients need to feel comfortable and trust that AI is appropriately involved in their care?
Mostly they need to have been told, and to know that a human is still accountable for what happens with the information and data. Very few patients object to the tool itself once that’s clear. What erodes trust is ambiguity, the sense that something is happening in the background that nobody bothered to explain. And honestly, the strongest signal is the clinician themselves. If the clinician is relaxed and can explain the tool in a sentence or two, the patient generally is relaxed too. If the clinician seems unsure about it, that’s what the patient picks up on.
If we move beyond efficiency as the main measure of success, how should we think about what ‘better care’ looks like in a system where AI is handling large parts of documentation and information capture?
Efficiency matters, and I wouldn’t want to undersell it, but it was always going to be the easiest thing to point to. It’s not the whole picture. What I look for is whether the record is more complete, because a clinician who isn’t splitting attention between the patient and the keyboard tends to capture a fuller picture, and the next clinician who reads that note is working with better information. I also look for whether clinicians leave a clinic with something left in the tank instead of 2 hours of notes waiting for them at home. That second one matters more than people give it credit for. A lot of the profession’s attrition sits quietly in the evenings, and if we can lighten their cognitive load, that’s what lets clinicians stay in medicine longer instead of burning out.
Healthcare professionals are often focused on whether a tool improves real clinical practice. Can you describe an example or piece of feedback that demonstrated to you that AI was making care delivery easier, safer, or more effective?
One of the clearest examples for me was a service in the NHS that had a backlog of over 2,700 clinic letters sitting unresolved, months of delayed correspondence that had been rated a ‘catastrophic’ clinical risk. After they brought Heidi in to support documentation, that backlog dropped to under 200 letters within 4 months, and the risk rating came down to ‘moderate’. In another Emergency Department, discharge letter turnaround went from over 9 days to 2.5 minutes. These are clear safety numbers: a stalled backlog means delayed follow-up, medication changes that don’t reach anyone in time and GPs working from information that’s weeks out of date.
The feedback that stuck with me most, though, was from a GP in a Primary Care Network in the NHS who described it this way: “I am no longer overwhelmed by the amount of information provided by my patients… As a result, I am able to fully focus on the patient and help them the best way I can.”
In healthcare, new tools often solve the problem they were designed for but create new friction elsewhere in the system. Where do you think AI documentation tools might unintentionally shift workload, attention, or responsibility in ways we’re not yet fully anticipating?
Responsibility has not fundamentally shifted, but the review burden has increased. Though the summary is generated, a clinician must still check it, and that check needs to happen while the clinician has proper attention to give, which isn’t always straight after a packed clinic. There’s also a slower risk around automation bias, where a clinician starts assuming the system has flagged everything worth flagging, when it was only ever built to summarise rather than interpret. Getting clinicians genuinely clear on where that boundary sits matters just as much as the tool’s accuracy. Training is really on us as suppliers to get right: there’s a genuine learning curve before a clinician can use these tools safely and effectively, and that takes longer to build than people assume. But once they’re over that curve, the unlock on documentation burden is vast, which is why I’d like to see this move toward curriculum integration rather than being left to induction week.
The EU AI Act is expected to significantly shape how AI systems are developed and deployed in healthcare. What real-world impact do you expect this regulation to have on how AI tools are adopted in healthcare systems?
Treating most clinical AI as high-risk means procurement teams will start asking harder questions much earlier: how was this validated, what data was used to train the model, where is the data held, how is the data protected, and what happens if there are AI-related errors? Whilst it may slow deployment down at first, it encourages healthcare organisations to be analytical, ask the right questions to tease out the vendors who take governance seriously, and encourages vendor excellence, because those who have already put in the work on structured validation and proper audit trails will find this straightforward. Vendors leaning on unsupported marketing claims will not. I’d also expect post-market surveillance to become a standard ask rather than a box ticked once at launch, and regulatory sandbox environments are a natural support to this, as they give vendors a supervised route to prove a tool works before it’s deployed at scale. This is a far more realistic standard than expecting everything to arrive perfect on day one.
Looking ahead, what do you think will change most in the use of clinical AI across European healthcare systems over the next 12 to 18 months?
I think the question shifts from whether to use AI clinically to how it’s governed. Procurement is likely to get sharper as a result, with health systems starting to expect something like an agent safety and efficacy profile up front and weaving ongoing performance into the procurement and pricing conversation itself, rather than treating monitoring as a one-off validation stamp. I would expect the dissolution of companies offering single-point tools that solve a discrete pain point, and rather a move toward companies that can support multiple aspects of AI-enabled care. What I think will take longer, and matters more, is the cultural shift: enabling clinicians with the training, knowledge, and confidence so that they are comfortable using and trusting AI systems, with the same habit of checking work as if for a junior colleague.
