Putting Patients at the Centre of Healthcare AI

There's a persistent assumption in healthcare AI that we already know who will and won't accept these technologies: that older patients won't engage, that people fundamentally want human contact, that automated systems inevitably feel cold and impersonal. These assumptions often go untested, and when they are tested, the findings are frequently more encouraging and more nuanced than expected.

At Ufonia, we've had the opportunity to collect systematic feedback across a large and diverse patient population, with over 110,000 satisfaction responses from more than 165,000 patient interactions across multiple NHS clinical pathways. What follows is a reflection on what measuring acceptability carefully has taught us, and how recent cataract surgery studies from Canada and the Netherlands demonstrate that these findings translate across different healthcare systems and patient populations.

Proving acceptability, not assuming it

If healthcare AI is going to earn trust, it has to be held to high standards, and that means not assuming acceptability, but measuring it rigorously. From the outset, we asked: Would patients want this? Would they trust it? What would put them off? Early clinical studies at Buckinghamshire Healthcare NHS Trust, UK established high levels of patient acceptability across multiple measures (Khavandi, Lim, Higham et al., Eye, 2022), and a subsequent NIHR-funded study across Oxford University Hospitals and Imperial College Healthcare demonstrated strong clinical agreement between AI and ophthalmologists for cataract follow-up consultations, with 96.5% of calls completed autonomously (Meinert, Milne-Ives, Lim, Higham et al., eClinicalMedicine, 2024).

But numbers only tell part of the story. Over 165,000 patient interactions across more than 22 NHS organisations later, we ask patients to rate their experience and tell us why at the end of every automated conversation. Over 110,000 of those responses now form a continuous feedback loop, helping us identify emerging themes, spot risks early, and understand what patients actually experience rather than what we assume they experience. Our most recent post-market surveillance data shows an 87.9% satisfaction rate and a Net Promoter Score of +56, placing us in the "excellent" category. That kind of systematic listening is something any team deploying AI in healthcare should build in from the start.

Recent international studies confirm these patterns hold beyond the UK. A Canadian study at surgical centres in Toronto found similarly positive results, with an NPS of +36 and strong usability scores for clarity, simplicity, and ease of learning (Hatamnejad, Higham, Somani et al., AJO International, 2026). Research from Maastricht tested the first Dutch-language version of the system, with NPS scores of +13.5 at week one and +12.6 at week four (Wanten, Bauer, Chowdhury, Higham et al., JMIR Formative Research, 2025). The Dutch scores were lower, partly because patients received automated calls alongside standard care rather than instead of it, and partly because the system used a translation layer rather than native Dutch language processing. That finding directly shaped our approach to multilingual development: native language capability matters.

Beyond the transactional

Some parts of healthcare are transactional: confirming appointment details, relaying clear instructions. AI can handle that well. But healthcare is also deeply human. Patients aren't just trying to complete a task. They're often trying to feel less anxious, less alone, and more in control of what's happening to them. Any AI system operating in this space needs to reckon with that. The goal shouldn't be to process patients more efficiently, but to support them in a way that feels respectful, clear, and safe. Recent linguistic research examining how automated clinical conversations actually function in practice is beginning to explore this territory (Brandt, Hazel, McKinnon et al., Discourse & Communication, 2024).

Listening to voices we don't always hear

If healthcare AI is going to work well for everyone, developers have to actively seek out the people who are most often left out of digital health design. Working with Health Innovation Oxford and Thames Valley, we've run focus groups with people from different language communities, people with lived experience of autism and learning disabilities, and others whose needs are typically missed when services are designed around "typical" users. It's easy to assume equity, and much harder to measure it. Focus groups with Turkish- and Polish-speaking communities, for example, have revealed how patients describe symptoms using culturally specific expressions, and how expectations around reassurance and escalation vary between groups. These aren't insights you can get from a translation service, and they should inform how any healthcare AI is designed.

The challenge we can't ignore: language barriers

One of the most underestimated challenges in healthcare is language. Clinicians know how easily meaning gets lost. Too often, communication depends on relatives interpreting, or patients nodding along without truly understanding what they've been told. In our focus groups, one message came through again and again: patients want to be able to describe their symptoms in their own words, in their own language, not through a family member, and not through an unfamiliar interpreter.

No clinician, however skilled, can be fluent in every language their patients speak. But multilingual AI has the potential to offer something genuinely new: consistent understanding at scale, across pathways, in the patient's own language. An NIHR-funded clinical trial, a partnership between Ufonia, Moorfields Eye Hospital, the University of York Centre for Assuring Autonomy, and Health Innovation Oxford and Thames Valley, is now testing what safe multilingual clinical AI looks like in practice. The trial involves 800 cataract surgery patients at Moorfields, extending automated conversations into ten additional languages including Polish, Bengali, and Spanish. This isn't just translation. It's the ability for patients to communicate naturally, in the language they're most comfortable in, and still receive safe, clinically appropriate support.

Designing with patients, not just for them

Patient and Public Involvement isn't a box to tick. It's how responsible healthcare AI gets built. Alongside patient feedback and clinical evidence, we've worked with the University of York's Centre for Assuring Autonomy to develop a structured ethics assurance case, applying a principles-based framework to reason systematically about the ethical acceptability of deploying an autonomous voice agent in clinical care. That process surfaced important considerations we hadn't fully anticipated, including risks to clinician autonomy when AI systems take on tasks previously performed by healthcare professionals (Kaas, Porter, Lim, Higham et al., TAS '23, 2023).

The lesson from all of this: patients don't need perfect technology. They need technology that listens. Building that means keeping patients at the centre, not just in design, but in how we measure success, who we include, and whether we're genuinely removing barriers to care.

---

Key references

• Khavandi S, Lim E, Higham A, et al. User-acceptability of an automated telephone call for post-operative follow-up after uncomplicated cataract surgery. Eye (2022)
• Meinert E, Milne-Ives M, Lim E, Higham A, et al. Accuracy and safety of an autonomous artificial intelligence clinical assistant conducting telemedicine follow-up assessment for cataract surgery. eClinicalMedicine (2024)
• Hatamnejad A, Higham A, Somani S, et al. Feasibility and Safety of an AI-Driven Postoperative Telephone System for Cataract Surgery Follow-up in Canada. AJO International (2026)
• Wanten JC, Bauer NJC, Chowdhury M, Higham A, et al. Optimizing the Postcataract Patient Journey Using AI-Driven Teleconsultation: Prospective Case Study. JMIR Formative Research (2025)
• Brandt A, Hazel S, McKinnon R, et al. Educating Dora: Teaching a conversational agent to talk. Discourse & Communication (2024)
• Kaas MHL, Porter Z, Lim E, Higham A, Khavandi S, Habli I. Ethics in conversation: Building an ethics assurance case for autonomous AI-enabled voice agents in healthcare. TAS '23 (2023)

Funded by the National Institute for Health and Care Research (NIHR) AI in Health and Care Award (AI_AWARD01852) and NIHR Invention for Innovation (i4i). The views expressed are those of the author and not necessarily those of the NHS, NIHR, or the Department of Health and Social Care.