Webinar Insights - AI-Powered Communication Skills Training for Healthcare Learners

At its core, SimFlow.ai is a speech-to-speech simulation platform. You speak directly to a voice character — an AI patient, relative, or colleague — and that character responds in real time, in a voice matched to their age, nationality, and gender. We layer ambient background sound over the conversation to heighten the realism. It's fully web-based and mobile-responsive, so learners can use it on any device, wherever they are. And we've just brought in SimCall, which takes simulations onto a real phone call — perfect for practising telemedicine and telephone consultations.

I've been practising medicine in the NHS for over a decade, and during that time I was teaching and training doctors, nurses, and undergraduates. What I consistently saw was this apprehension — trainees were nervous about having challenging conversations with patients and relatives. And even when the confidence was there, the opportunity often wasn't. Sensitive conversations would happen behind closed doors. Students would be excluded entirely. Traditional simulation with actors and simulated patients is still the gold standard. But it's not scalable and it's rarely available. SimFlow.ai was born out of that problem — a real gap we were seeing clinically and educationally. That was three or four years ago. Now it's being used by NHS trusts and universities internationally.

SimFlow.ai is designed to sit alongside and enhance in-person teaching, not replace it. Actor-based simulation remains the gold standard. But government and NHS workforce plans are pushing to significantly increase student numbers in medicine and nursing, and the resources — facilitators, actors, physical space, time — are already stretched. SimFlow.ai fills those gaps.

The thing I keep coming back to is that it's asynchronous. Learners don't have to wait for a facilitator, another learner, or a patient to be available. It's there whenever they need it. I used this with my GP trainees preparing for their final-year exam — I created a group within the platform, allocated them cases, and left them to work through these independently. All of them fed back that it was genuinely useful.

Technically, we could implement video avatars within a week. It's not a capability question. But we've been really clear from the start, informed by our clinical background and our academic partnerships, that we wouldn't do it until the technology is genuinely ready. If I deliver a difficult piece of bad news, the avatar's facial expression isn't going to change significantly enough to match the weight of that conversation. When the visual and the emotional content are incongruent, you've actually created a worse learning experience, not a better one. It's not a question of if — it's a question of when.

This comes down to cognitive load. If you've got a face that isn't representing the speech — someone's deeply upset but the avatar looks blank — your brain has to process conflicting signals. That's extraneous cognitive load, and it eats into the capacity you have for actual learning. What learners need is something functionally good enough to mirror real activity. That's what SimFlow.ai's voice-based approach achieves right now.

We started building an evidence base from day one. Four years on, we've progressed from quality improvement projects through to multi-site clinical research with a range of university partners. Published findings cover use in medical schools, the armed forces, and the secondary school sector for teacher training. SimFlow.ai's published cost-benefit analysis demonstrates an 84% reduction in cost compared to actor-based training — without compromising on fidelity or acceptability as a simulation tool.

What drew me to SimFlow.ai in the first place was that when you ask for the evidence, there is evidence. As a health technology researcher at the University of Bath, that's genuinely rare in the current AI landscape. Most developers, when I ask where their evidence is, I get a blank face. Jon said, here it is. I love that. Technology in education should be open to scrutiny.

We have two simulation modes. Realtime is a free-flowing, fluid conversation that closely mirrors real clinical interaction — best suited to more experienced learners. SimVoice, our original model, is turn-based: you hold down the spacebar to speak, then wait for a response. It lets learners slow the pace right down, which is really valuable for first and second year students who are still building confidence.

And the asynchronous element matters enormously here. People practise at all different times. Learners don't have to wait for me to facilitate. They can pick it up at 11pm the night before an exam if that's when it suits them.

The mastery that comes from repeated, structured practice is well established for physical procedural skills. We all accept that surgeons need to practise suturing. But we apply that same logic inconsistently to communication skills, even though the evidence says it works exactly the same way. SimFlow.ai gives learners a vehicle for that deliberate practice — they can repeat scenarios, target specific areas of weakness, and track their development over time. And it taps into intrinsic motivation. Learners want to improve. I use SimFlow.ai myself as a practising GP. I had a case recently where I felt my mental health risk assessment wasn't thorough enough, so I went into the psychiatry scenarios and worked through a few. That's exactly the kind of self-directed, targeted practice it enables.

The library covers psychiatry, medicine, dentistry, pharmacy, nursing, midwifery, clinical psychology, child and adolescent health, pre-hospital care, the armed forces, police, and NHS governance and appraisal. We also cover more specific situations: language barrier navigation, breaking bad news, deprescribing, smoking cessation, telephone appointments, vaccine hesitancy conversations. And critically, it's not limited to patient interactions — scenarios include conversations with relatives, carers, colleagues, and MDT members. Organisations can also commission entirely bespoke closed libraries. We build those collaboratively, vet them, release them, and iterate on them over time. And one thing we didn't fully highlight in the demo is that we can follow the same character across time — you might meet the same patient as a teenager, then in their thirties, then later in life. That cradle-to-grave approach to longitudinal training simply isn't possible with actors.

Assessment is built into every simulation. Before each scenario, learners can complete an optional pre-brief — reflecting on their existing knowledge, confidence, and what they want to develop. After the simulation they answer post-brief questions. Then an AI-generated report provides structured feedback across twelve domains: six for information gathering, six for communication skills. Rather than just giving scores, the report pulls direct quotes from the conversation as evidence for the feedback it offers, and makes specific suggestions about how a response could have been framed more effectively. It also cross-references the pre-brief: if a learner said they felt unconfident in a particular area but actually performed well, the system will notice and comment on that discrepancy. Learners get a full transcript, audio playback, and can download a PDF of the report for their portfolio.

Absolutely — and it's something we consider a core part of the platform's value. When we build custom scenarios with an organisation, the assessment report is built alongside the scenario. Getting the character to behave correctly is only half the task. The other half is making sure the assessment report responds to what that organisation actually wants to measure, and at the level appropriate for their learners. For the open library, the assessment is intentionally broader to accommodate the full range of users — from a student who's a couple of weeks into a placement, all the way to a qualified professional — but it's still designed to pick up on the key elements relevant to each scenario's learning objectives.

We used it very successfully with GP trainees at ST3 level — the final year of GP training — who face an OSCE-style assessment conducted via video call. I created a cohort group within the platform, allocated scenarios tailored to the exam format, and let the trainees work through them asynchronously in the run-up to the exam. All of them reported finding it highly useful. The standardisation benefit is significant. Because the AI generates consistent scenarios, every learner is practising against the same content — that's extremely difficult to guarantee with role play or actor-based preparation, and it matters a lot in high-stakes exam contexts.

I've flipped my teaching sessions entirely with SimFlow.ai. Rather than delivering content in a lecture and then discussing it, I assign the pre-reading and the scenario as preparation before the session, then bring students in to practise the consultation skills live. It's significantly more effective than the traditional format — and more enjoyable for the learners. The built-in pre-reading, electronic patient records, and pre-brief prompts support structured independent preparation, while the post-brief questions and assessment reports give me meaningful material to draw on when students arrive for face-to-face teaching. Education should be fun, and I think that's part of why this works.

One of SimFlow.ai's distinguishing qualities is its usability. As a GP educator who has integrated SimFlow.ai across several teaching contexts, I assessed it using the System Usability Scale — a validated tool for measuring how easy technology is to use — and the platform achieved a Grade A score. That's genuinely difficult to achieve. A lot of AI tools in education require a technical person to be present to support users, which immediately creates a barrier to scalability. SimFlow.ai doesn't. In terms of curriculum integration, it does require some initial conversations within departments. But the process is relatively straightforward because the cost isn't prohibitive. When you present a clear business case — cost savings, improved student performance, high satisfaction scores — it's a compelling investment.

It's primarily designed as a screen and phone-based platform rather than for use with physical simulation equipment, but a number of people have found practical ways to make it work. The most common approach is connecting it through a Bluetooth speaker placed near the manikin's head, and the feedback we've had from that has been really positive. The main consideration is background noise — in a busy simulation suite, ambient sound can feed into the microphone and either disrupt the AI patient's responses or prevent them from triggering. For smaller, quieter scenarios this works well, and you can also turn off the platform's ambient sound feature to help manage this. The majority of our 400 to 500 scenarios don't have ambient sound enabled by default.

SimCall — where any scenario can be conducted as a real phone call made directly through the platform — is coming in as a third simulation mode. Some of the early anecdotal feedback already is that students on phone-based simulations haven't realised they were talking to an AI patient. They expect a tiny bit of lag on a call, so it felt completely natural. That speaks to the level of immersion the format can achieve.

When we started, we leaned on the wealth of clinical experience our team had — a decade of encountering all kinds of cases across different specialities. But as the libraries have grown, we've moved towards a partnership model with the trusts and universities we work with. They bring forward actors, expert patients, and case studies, which we then anonymise and build into the platform. We've done that in both healthcare and the school education sector.

We also involve different clinicians in writing scenarios. We've had surgeons contribute surgical cases at Great Western, for instance. That clinical credibility in the design is important — you need people who know what these conversations actually look and feel like.

If we wanted to implement video avatars tomorrow, we could. What we've seen with current technology is that it doesn't appropriately match the emotional content of a conversation. The avatar can move its head, maybe its hands — but if I break a difficult piece of bad news, the facial expression isn't going to change significantly enough to match the weight of that moment. That incongruence undermines the educational value. It's not a question of if, it's a question of when — and we'll make that call when the technology is ready.

On ambient sound — it's worth noting that 90 to 95% of the scenarios on the platform don't have ambient background sound enabled by default. We used it today for demonstration purposes. So the vast majority of learners will start without it, and there's scope to think about switching it on as a deliberate curricular decision as learners progress. On manikins — we've seen a lot of people connect SimFlow.ai through a Bluetooth speaker placed near the manikin's head, and the feedback has been positive. The main thing to be aware of is background noise in a busy simulation suite, which can feed into the microphone. For smaller, quieter scenarios it works really well.

Traditionally we work around a simulation credit model — each conversation is one credit, and conversations can run upwards of 20 to 30 minutes if needed. There are also seat-based pricing and wider licence options. The best approach is always to have a conversation with us directly, because the right package really depends on your cohort size and how you plan to integrate the platform.

This was something we were really keen to get right from the start. When we build custom scenarios with an organisation, the assessment report is built alongside them — aligned to their rubrics, their curricular objectives, and their learner standards. Getting the character to behave correctly is only half the work. The other half is making sure the assessment responds to what that specific organisation wants to measure. For the open library, there's naturally a bit more breadth to the assessment — it needs to work for everyone from a student two weeks into a placement to a qualified practitioner. But it's still designed to pick up on the key elements relevant to each scenario's stated learning objectives.