Episode 1:Simply Onno

Welcome: who is Simply Onno?

Andreas Dahrendorf: Karen Hentschel and Marc Dantas Tiedemann founded Simply Onno, a service that uses AI to explain doctors’ letters and medical findings in plain language. Karen was Design Director at IDEO, and Marc helped build Heartbeat Medical. In 2025, Simply Onno won the WELT German AI Award as Best AI Start-up. A very warm welcome, Karen and Marc.

Karen Hentschel: Hello.

Marc Dantas Tiedemann: Hi, great to be here.

Andreas Dahrendorf: To start us off, the first question: there are four founders in three locations. Who does what?

Karen Hentschel: We’re actually in four locations now: I’m in Munich, Marc is in Berlin, Witold, our doctor, is in Dresden, and Stephan is in New Zealand. We all have different areas of responsibility, of course. Since I’ve already started speaking, I’ll just say what I do.

You’re right that my most recent job was at IDEO, a design consultancy. My background is in design. That’s also how Marc and I know each other: we studied design together in Potsdam. Afterwards, I did a lot of product design and also built design teams. So I mainly look after design and strategy here, and everything to do with user research, meaning what we can learn from the people who use or may use our products in the future.

Marc, I’ll hand over to you in a moment, but first a little about Witold and Stephan. Witold works in neurosurgery at Uniklinikum Dresden. He takes care of everything medically relevant for us, from quality assurance to checking that the output is correct. He’s also fantastic on regulation and advises us on that. Then there’s Stephan, our CTO. He has a lot of experience building and developing digital products and complex platforms, and understands what it takes to create services in our context.

Marc Dantas Tiedemann: My role is a bit of everything. There are only four of us, so there’s a lot to do across the board. That ranges from marketing, which overlaps a lot with Karen’s area, to product development. I also do a lot of prompting and try to keep all the threads together.

As you mentioned, before this I was at Heartbeat Medical, another digital health start-up. It scaled up quite a bit, and by the end we had up to 50 employees. Though, by the end: the company is still going. I left after ten years because I wanted to try something new and smaller. Heartbeat worked with hospitals, with all the problems that come with that, such as selling to hospitals. I wanted to do something smaller and faster, directly for patients, something you can launch and that has an immediate impact.

Andreas Dahrendorf: Onno’s founding story begins with a medical report of your own, Marc, which you could read but didn’t understand. Can you tell us about that?

Marc Dantas Tiedemann: A few things came together. I tore something in my foot while bouldering, and I also had recurring knee problems for a while. So I had a few MRI reports lying around.

Really, though, it started even earlier, or around the same time: Karen and I were both available at the same time. As I said, we’ve known each other a long time and have been friends for years. We go snowboarding together every year and talk a lot while we’re there. Something we haven’t mentioned yet is that we’d already done some pretty successful, cool projects together at university. For a long time we kept saying we should do something together one day, and then suddenly the moment arrived.

Originally, we thought we might do something together in design leadership. We considered job-sharing, applying for a role together, or starting an agency. Then, completely off the cuff, we said: we haven’t worked together for ten years, so we should test it first. Does it still work? We had a rough idea, and the topic of medical findings was already on our minds. We said: let’s just try out how we would do it. The rest is history. We decided to stop applying for jobs, see the idea through and start our own company.

Andreas Dahrendorf: What happens when I upload a doctor’s letter?

Karen Hentschel: Most people are holding a medical document, a test result, a doctor’s letter, a CT or MRI report. There are so many complicated documents. That’s how it was for me at the time, too. They contain a lot of complex information, and you’re unsure what it means or whether you’ve understood the doctor correctly. So you Google the terms, end up on one of our many pages explaining terms, and can get your report explained in full there.

In practice, you say: I’d like an explanation of my medical report. Then you either take a photo with your phone or upload the file on your computer. The document is shown to you again, and you can confirm that it’s the one you want explained. Then you ask for an explanation. It takes a little while, and while that’s happening we let you know what’s going on in the background. It can’t be done in two seconds: we prepare the explanation and check that it’s medically accurate and meets our quality standards before displaying it. The result appears on the device you used to upload the document.

What you receive depends on what you’ve chosen, because we have two products. One is the free summary. It now goes into quite a lot of medical depth. We’ve iterated on it several times to make it better and meet users’ needs more fully. It clearly explains what’s in your document: What’s the diagnosis? What is it saying, broadly speaking? It’s free, and you get it within a few seconds.

Or you can choose the paid version, which has been available for a year now: a detailed explanation for a one-off fee of EUR 18.50 per document. It takes a little longer to prepare, but then you get a report of several pages that explains every single word and describes and contextualises the diagnosis. It also covers possible next medical steps, what it might mean, and what you might be able to do yourself with your diet or exercise. And, very importantly, which questions you should ask the next time you see your doctor.

I say “very important” because that is the most important thing for the people who use us: knowing what is still unclear. Those questions usually only occur to you after the appointment. You may know that feeling from your own visits to the doctor.

To answer your question: broadly speaking, you upload your document and, depending on which option you choose, you get a more or less detailed explanation sooner or later. You can save it as a PDF. You don’t have to create an account; we don’t know who you are. The only thing you give us is your document. That’s a deliberate choice: we want to make the barriers as low as possible so that everyone can use it quickly, while giving people confidence that anything they share with us is handled in line with data protection rules.

Karen’s path to Simply Onno

Andreas Dahrendorf: Before we go further into the product, a quick question about you, Karen. Perhaps this is a link between the two: you previously worked on projects for refugees. Did anything you learned there find its way into Onno?

Karen Hentschel: Yes, it did. For context, while I was at IDEO I worked with Alight, an NGO whose CEO used to be the CEO of IDEO.org. IDEO.org is IDEO’s non-profit arm and does exactly this kind of work. Alight got in touch when the war in Ukraine began and asked for support. I agreed to work on the project.

That was certainly one of the reasons I came to think, over the course of my career, that what you do should also benefit everyone else. That’s why we’re now doing B2C, whereas Marc previously worked in B2B.

The thing I took away most from the project was trauma research: how do you interview people who have been severely traumatised or are living in that kind of context, and how do you talk to them? That helped me enormously. In the research and in our work with Onno’s users, I meet people who have received serious diagnoses and talk with them to learn about our product. How do you do that respectfully, so every person you speak to feels equally included, and without causing further harm or putting any additional burden on them? The awareness of the context you’re entering was what helped me most.

Andreas Dahrendorf: Related to that is the question: when is an explanation accurate and simple at the same time? How do you do that?

Marc Dantas Tiedemann: It’s a topic we have big debates about again and again. Well, “debates” is putting it too strongly, but it’s a balancing act between medical accuracy and a simple explanation. It takes a lot of fine-tuning and repeated reviews. In the end, part of it is also gut feeling, at least when it comes to tone and simplicity. There are parameters you can use to calculate it, of course, but ultimately you have to decide for yourself. There’s a recurring tension between simplicity on one side and a doctor on the other saying: this isn’t accurate enough. You have to find the balance between them.

Karen Hentschel: We’ve also worked hard on that. The summary in particular went through many iterations, from the very first, fairly lightweight version to people’s need for more medical depth, while keeping the explanations simple enough for everyone to understand. Which medical terms are generally known, and which ones do you need to explain? Does everyone even know what an MRI is? We’d rather explain something one more time than lose someone.

That’s definitely a big part of our work. Medical quality is one part, but a huge amount of the work goes into how we communicate things and how complex or simple we make them. And to return to the situation in which people read our explanations: how do we need to communicate so that they can process information they may be hearing in this form for the first time, in the worst case, without us causing harm?

Just one more thing: when we created the English version, we found that it was a whole different challenge because language becomes complex much more quickly in English. Thankfully, there are fantastic public services in England that are very design-led and have given a lot of thought to how language should be used in contexts like these. They offer useful guidance. But you still have to ask: who is our audience, who needs to understand this, how complex can it be? Then you give the text to someone to read and ask: do you understand what’s in there now? Of course, we also get feedback from our users.

Andreas Dahrendorf: That was going to be my next question: feedback. How do you know that people understand, are satisfied and trust you? It is still AI, after all.

Marc Dantas Tiedemann: Mainly from the feedback. If you know Slack, it’s a communication tool where you can create different channels. Our feedback goes into a channel we call the Good Mood Channel, because it’s almost entirely positive. How does it work? At the end of an explained report, people can give it a thumbs-up or thumbs-down, as you know, and write a comment too. Quite a few people do that, I’d say around five a day.

It ranges from short comments like “Thanks, brilliant” to really detailed stories. Some are quite moving because you can see that we’ve genuinely helped at a critical moment, brought clarity or perhaps simply eased someone’s fear. That’s how we can tell it worked and that we got the tone right.

Of course, every now and then there’s a negative comment, such as “I would have liked more detail” or “I already understood that.” That came up mainly with the short summary. Depending on how often that sort of thing happens, you can see that you need to adjust it again. In all that time, I think there have been one or two cases where there were still minor errors. We’ve now explained 100,000 documents, so I think that’s still a pretty good record. The feedback also gives us important input on those cases, of course.

Karen Hentschel: There are two ways we do that. First, after the summary and the detailed report, we make it very easy for people to tell us whether what we’ve created was helpful. They can also select “Yes, because …” or “No, because …” and leave their email address. We contact the people who leave us their email and ask them to speak with us, or send them a more detailed questionnaire in which they can rate different things.

Karen Hentschel: That’s how we try to speak with users regularly. One reason is to understand whether they’re satisfied. It’s also an additional layer of quality assurance: we’re not creating new information, and we need to make sure everyone knows they should discuss it with their doctor.

On the other hand, this also shows us very clearly how we should develop the product. What do people actually do with the information? Where do they need support? These conversations let us find out in much more detail what works and what doesn’t, what people need the information for and how complex their circumstances are. Is this someone who has knocked their knee and simply doesn’t want to wait three months for their GP to explain the MRI? Or is it someone with a longer medical history who now wants to play a more active part and understand whether what they’ve been told will help them make their next decision?

Andreas Dahrendorf: Without breaching anyone’s privacy, can you describe who books and pays for it? Are they particular groups or typical users?

Karen Hentschel: Do you want to take that, Marc?

Marc Dantas Tiedemann: I can. We can generalise without naming names. At the moment they’re mainly older people, I’d say. Older meaning around 60 or 65. As Karen said, we send out these questionnaires, and based on the responses, I think the average age was 67. The oldest person was, Karen, remind me, 92 or 95?

Karen Hentschel: 94.

Marc Dantas Tiedemann: Right, 94. So the range is very wide, but it skews older. I think that’s also because people tend to have more medical findings as they get older. When I talk to friends, many of them haven’t had many test results or major doctor’s appointments yet. Then, after a certain age, it starts happening rapidly.

Andreas Dahrendorf: Sure.

Marc Dantas Tiedemann: And we’ve made our interface as simple as possible. No login, no app download: I just press this button and can upload something. It’s easy for everyone to use.

Karen Hentschel: To be a little more specific, there are different groups. First, there are people who need answers urgently. Like Marc with his foot: it hurt, he was referred for an MRI, had the scan, and then had to wait again until his GP had time to explain it to him.

Marc Dantas Tiedemann: And before that, I wanted to go snowboarding one more time.

Karen Hentschel: I have a friend who broke her thumb. It’s healed by now, but not as it should have, because it took so long for her to get the MRI and then receive an explanation. Those are people who need answers urgently, and they’re naturally more willing to pay. They say: I can wait three weeks, or I can have this translated and at least know what’s going on. Then I can arrange physiotherapy myself or put a bit of pressure on the system.

Then there are people who’ve just been diagnosed, as I’ll call them, with complex conditions, who say: I want to understand this in depth now. Doctor’s appointments are very short, and people are actually expected to deal with their diagnoses very independently as soon as they learn about them, to know what it means and what’s coming. There’s a need to understand the issue in depth and in its entirety.

And the more diagnoses someone has, increasingly chronic conditions with recurring information, the more they want to manage their situation actively. They hope to get information that helps them make the next decision, have the next conversation more effectively, know what to ask or put information into context over a period of time. We know from interviews too that some people pay more than once for different documents. So there are people who don’t just use it once and never come back. We do have users who repeatedly ask us for detailed explanations of new documents.

Free and paid: the model

Andreas Dahrendorf: You said you have a free product and one for EUR 18.50. We don’t want to get too deep into profitability, but what needs to happen for enough people to pay?

Marc Dantas Tiedemann: The business isn’t profitable yet. We’re still “bootstrapped”, as they say in English. That means we don’t currently have any investors on board. We’ve funded everything ourselves and aren’t paying ourselves salaries yet. It already covers the servers, software licences and other running costs, but it doesn’t cover salaries. So we’re not profitable yet.

We’ve also said that the free version should always stay free. For one thing, there’s a marketing aspect to that: you can recommend it to anyone with a clear conscience. We’re now talking to various patient advocacy groups, foundations, hospitals and the first insurers, and they can only recommend us because the free version exists. We also want to offer something with a low barrier to entry that as many people as possible can benefit from. Karen and I, and the whole team, are completely aligned on that. And the more widely it spreads, the more often we’ll also get a purchase, a detailed translation, of course.

We don’t want to stop there, naturally. We have lots of plans for how to expand it. We’re currently building an app so that it’s about more than translating a single report. That’s always very specific to one moment: in business terms, you basically have to acquire each person all over again. That’s hard to sustain over time and takes a lot of effort.

An app is much more interesting in terms of what it can do, too. Say I’ve injured my knee five times and had it operated on a few times. That was a few years ago, and now it’s hurting a little again, and I want to know: what exactly happened back then? Then I have to dig out the records again. If I had them in an app, I could see exactly: right, the physiotherapist told me to do these exercises, and I stopped doing them. We could then build other things around that. We’re thinking about a subscription model, and perhaps we could involve other partners too. There are lots of ideas for how to develop this further until it really becomes profitable for us as well.

Andreas Dahrendorf: Is the electronic patient record relevant to you in any way?

Karen Hentschel: Definitely. Above all, it reinforces why our service is needed. If you think about where we started, we had an individual need to understand this information. But we soon realised that we weren’t the only ones being left to work it out alone. So we wanted to use our skills to create a solution.

And since we started, the situation has become more difficult rather than easier. The electronic patient record (ePA) is now available, which means even more people can access this kind of information. Previously, it was often sent directly to doctors. Now people can access it themselves. At the same time, the legal situation is changing in a way that makes it increasingly difficult to speak to specialists about it. So people are being left more and more to manage on their own. That’s also why we said we need the free product: people shouldn’t have to be able to afford healthcare in order to understand it. We want everyone to be able to understand information they already have so they can take an active part.

That’s why the ePA is a huge topic for us. We’re also talking to health insurers, which can see there’s a need. They’re right at the interface and could make this information understandable for their members, who would then be much more likely to follow their treatment plans or assess what’s happening.

We’re always looking for figures that show why this matters. There are some striking statistics, like, I think, three-quarters of the population have low health literacy. That means they don’t understand this kind of information at all. Or, I think, half of the people leaving hospital haven’t understood the diagnosis that was explained to them. That’s astonishing. And the more access people have to these documents, the greater their need to understand them is likely to become. If I don’t have them, it’s perhaps different. But if I’m discharged with this information and am now essentially responsible for my own care, of course I want to know what it means.

Andreas Dahrendorf: I assume a lot of people also use other AI tools. Have you asked how your results and answers differ? Privacy must be a factor compared with open, general-purpose AI tools.

Marc Dantas Tiedemann: Privacy is obviously a huge topic. Everything is stored on European servers, in fact only in Germany, and the processing also takes place in Europe. We’re very transparent about it and say: ideally, cover your name too. We don’t want to see it. Even so, we check the documents again in the background to make absolutely sure nothing is left in them when we review them. Whether the big American companies do the same is always a question.

But to be completely honest, in our experience we think a lot about this, while users out there usually aren’t that concerned about it. People actually send us a lot of things by email, and we always say: please upload them on the website. We don’t want to receive them, especially not with your name and story attached. People out there are much more open about it.

Then, of course, there’s medical accuracy. We like to compare it to a Swiss Army knife. An LLM like ChatGPT can do all sorts of things. But would I want a surgeon to operate on me with a Swiss Army knife? Probably not. I’d rather have a scalpel, a tool that does just one thing, but does it very well. That’s a very visual way of putting it.

Since we only translate and explain medical reports, we can focus entirely on that. We go a little further than a straight translation and explain some things as well, while staying very close to the context: what did the doctor actually say? We haven’t trained our own model. We work with the large models available, but we’ve narrowed their scope considerably. Other AI systems evaluate the results, and doctors review them to make sure they stay within those bounds. We make a very clear distinction between what the doctor said, what the letter says, and, when we go a little further, what constitutes general information. Those are always clearly separated. Any further information is based only on guidelines, studies and sources specific to Germany. There’s quite a bit more to it, but that’s the core of it.

Karen Hentschel: Of course, we’re designers, so perhaps we have a particular eye for this: a chatbot isn’t the right interface for everyone. And quite apart from that, our users, whose age group we’ve already discussed, may not yet be used to interacting with chatbots. That will certainly change. But different needs call for different interfaces. With a chatbot, you’d need to know exactly what you want it to explain and in what form to get the result you want.

We did a lot of research and thought carefully about what information our users need at that moment, how it should be presented and in what format. There’s a reason it’s a PDF, for example: people want to print it out, file it and take it to their doctors. There are a lot of design decisions that are really important and shape how the product gets used. You can’t simply replicate them with the chatbots out there.

There’s also the trust factor, and what Marc mentioned earlier: doctors and foundations recommend us, saying that because we meet their quality standards, they can recommend us with a clear conscience. They’re grateful when patients don’t come in saying, “ChatGPT said I should …”, leaving them to untangle that in the little time they have. Instead, patients bring something they can build on, because they know the foundations are sound.

This is especially exciting because a study was published just yesterday in the New England Journal of Medicine AI that demonstrated exactly this. At Dresden University Hospital, they used our summaries in neurology appointments and investigated whether communication between patients and doctors is affected when patients read one of our summaries beforehand. They randomly assigned 300 patients who had read one of our summaries and compared them with 300 who hadn’t, then looked at what actually happened in those conversations.

The findings were that patients felt better able to take part and make better decisions, so the overall quality of the conversation changed. It’s really important to create that kind of foundation so we can work in this context and be recognised by all the people whose involvement is needed to handle a situation like this.

Andreas Dahrendorf: That’s fascinating. Do you want to say more about it, or have you taken part in other studies as well?

Marc Dantas Tiedemann: No, that was the first one. I think there were two parts, one in an inpatient setting and one in an outpatient setting. I think it was the outpatient study that was published now.

Karen Hentschel: This is really great for us. We can make all sorts of claims, but it’s fascinating to examine them in a clinical setting: are our assumptions right? Is what users tell us right? And then seeing what effect it has is fantastic. It was a small group in a very specific context, and neurology is very complex. So we can’t assume the findings apply in the same way everywhere. But I think we can say that summaries and clear explanations of complex medical documents help patients have better conversations with their doctors.

We all know the feeling: you go to the doctor, they explain what’s in the report, and there’s very little time. In the worst case, you have to process what you’ve just heard first. Then you go home, all the questions occur to you, and you start trying to make sense of it. If you’ve had time to understand it beforehand, you can go into an appointment much more prepared and ask: what are we actually talking about here?

This helps us too. Doctors were already recommending us to their patients. But the study also found that the conversations were shorter, while patients still felt more involved. That’s an incentive in the real world: if we can help patients feel better prepared in less time and give them more, everyone benefits. That’s great for us and really exciting.

Marc Dantas Tiedemann: That also fits with what we heard in our initial research, when we interviewed a lot of doctors. While we were still developing the product, we thought doctors would tear us to shreds and say: no, this is our area, keep your hands off, you’re not going to explain anything to them. It was exactly the opposite. We spent a week talking to experts about what they thought, and every single one said: oh my God, it’s great that you’re doing this! We want to explain things, but we don’t have time to go into them in depth.

I think the average doctor has seven minutes. The study dealt with more complex cases, and I think they had 20 or 30 minutes in total. Usually, though, appointments are quite short. Doctors have to keep explaining the basics, the same things over and over, which is tiring, of course. Then patients forget half of it and ask again. Having the space and time for the really important questions adds enormous value.

Karen Hentschel: Another thing we shouldn’t forget is that the patient often isn’t on their own. There may be family carers who handle the communication but aren’t necessarily there when this information is shared. Or there may be family members involved in making decisions. It’s great if they can all have the conversation from the same starting point. Or I might go from one specialist to another without having to explain the whole story from scratch every time, because we have a shared understanding. We shouldn’t underestimate that. Perhaps because of our users’ average age, but in general, especially with more complex diagnoses, this often affects more than the patient themselves.

Technology, AI and data protection

Andreas Dahrendorf: Let’s look at the technology: what AI is behind it? You’ve said it’s hosted in Europe and Germany. Can you give us some details about what works?

Marc Dantas Tiedemann: You can actually read all of this in our privacy policy, so I’m not revealing any secrets. Ultimately, we use the large models that are currently available. We’re a small start-up, and cost is one reason for that.

Karen Hentschel: And quality is another.

Marc Dantas Tiedemann: Exactly, both. A document goes through several stages: anonymisation, classification, then translation and a review process. Some of it runs on OpenAI’s models, hosted by Microsoft in Europe so that the data stays here. We also use Anthropic, but only with anonymised, redacted text, because Anthropic simply has the best models, especially for language. We’ve tested them against one another extensively, and that was the best combination. We evaluate it continuously, of course.

We’re already working on what comes next: new open-source models keep coming out and getting better. We’re actively looking at which parts we might be able to host entirely ourselves after all. That brings other cost factors, from maintenance to licensing. But we’re already looking at how to move everything over gradually.

Andreas Dahrendorf: Have you also looked at Mistral? Have you got that far yet?

Marc Dantas Tiedemann: I’d have to ask Stephan. I don’t think Mistral performed very well, unfortunately. It’s more the Qwen models and similar ones. They usually manage the medical precision. The tricky part is getting the text itself to sound natural as well. When I look at it through my own lens, I can tell if something is awkwardly phrased, grammatically wrong or just doesn’t sound right. That’s one of our main USPs. It’s the product, after all. So we’d have to make some serious compromises there.

Andreas Dahrendorf: And what does quality control look like after the cascade you mentioned, to make sure nothing is explained incorrectly?

Marc Dantas Tiedemann: There’s actually a whole cascade before that. Before we put anything into production, we have major evaluation and iteration rounds. To put it very simply, you come up with a system or prompt for one area, implement it and test it on perhaps ten reports. If, for example, Witold says he’s happy and we’ve solved that particular problem, then it goes into the big evaluation, which we all dread a little because it’s so time-consuming. We run 50 mixed reports, including lots of edge cases, and people actually assess them.

We’ve set criteria for that: at least one doctor has to be involved, along with at least one of us. I think we’ve now decided to have two people looking at readability. We have different scores: Medical Correctness and Compliance, which only doctors answer. Then Readability, meaning whether everything has been translated. I think we also had Helpfulness as a score, along with form and structure.

Karen Hentschel: And Personality and Empathy, things like that. Those are the criteria we test against.

Marc Dantas Tiedemann: Exactly. Then it’s assessed, and if anything fails, it has to be done again until it reaches a certain standard in that batch at least. We also run what’s called LLM-as-a-Judge on top of that: other, smaller models trained specifically on these scores that assess the whole thing again. We’ve found that sometimes they’re even better than us humans, not in every case, but because they’re simply more objective.

Sometimes I think I just want to get this finished. Or I judge it differently depending on the time of day when I’m reading it. That definitely happens. Or it depends on what conversation you’ve just had about certain topics, and whether you’re paying closer attention. A model doesn’t have any of that. It judges things fairly objectively and sometimes very bluntly, and then you think: oh, damn, it’s right. We combine the results until an evaluation passes and we can say: the scores from both the people and the machines are all great, so we can release it.

In the live system, it depends on the stage. Right at the beginning, when we went live, we really did look at every report afterwards. We don’t do that any more. We now have over 200 translations a day, so we check samples. We also have these LLM judges running in the live system. They score and assess the results and can intervene: if a translation falls below certain thresholds and still doesn’t pass on a second attempt, the process is stopped and nothing is delivered. Then we can investigate exactly what went wrong.

I can tell you a real-life example: recently we had a case where the text addressed a deceased person directly as “you”, and somewhere it also said “fortunately”. That’s obviously a problem. We have to intervene and refine the system. We can monitor all the scores for that.

Karen Hentschel: And there’s another very simple step before that: we check what kind of document it is so certain content never reaches us in the first place. Is it really a medical document, or something else? We also don’t translate every document. Our guiding principle is: you’re anonymous, we have no accounts, and we know nothing about you. We explain, we don’t create new content. That also means, for example, that we can’t really explain blood test results, because they lack context. Without knowing certain things about the person, you can’t say how to interpret a value. So for now, we don’t handle them.

If we eventually have an app with accounts, we can consider what information it would need. Our goal isn’t to replace doctors. It’s to give people the ability to make the right decisions and the tools they need to take action themselves, whatever form that takes.

Andreas Dahrendorf: Looking back at errors that appeared later in the product, or were at least caught, meaning in operation rather than during training, were they more technical errors, basic things like something being unreadable? Or were they already more complex questions of content?

Marc Dantas Tiedemann: Actually, both. Sometimes something wasn’t translated in full detail. Or there was the example I just mentioned, where the text suddenly addressed a deceased person as “you”. That’s somewhere in between, a question of understanding the content. We’ve had things like that with children too: children were addressed as “you”, even though they’re not officially allowed to upload reports, and it really makes no sense for a three-year-old. There are stories like that, and sometimes issues of medical accuracy that you have to look into closely. It took a long time and several rounds of iteration, and a few things also came up in production. They’ve all been fixed now.

But there were still moments when the model added a judgement, saying “fortunately” or “oh, it’s not that bad”. Or what ChatGPT and all those models love to do: offer reassurance, saying “it’s only something minor, it’s due to your age”. Things like that kept creeping into the content, and we had to say: no, we need to refine this.

Karen Hentschel: That’s like any product, just more critical because we’re dealing with medical information and reports. Otherwise, we’re a digital product developed iteratively. There are things we hadn’t thought of or had on our radar until they came up, and then we refine things. From a medical perspective, we’ve been compliant from the outset. That’s why we do this evaluation, and we only release things that are accurate according to it.

Then there are nuances like: oh, wait a minute. We said we want to feel close to people, so we address them informally as “you”. But that doesn’t fit certain contexts, and we have to find a solution for that. When something like that comes up, we know we need to think it through and tackle it straight away. At the beginning, we also thought we were a medical device and had that assessed. That’s why we develop the product this way too: all our quality standards are as if we were a medical device. That’s the foundation underneath it.

What happens after that tends to be smaller things, but they matter a lot to us personally because we hold ourselves and our product to a standard we want to meet. So those things aren’t acceptable, and that’s why we keep working on them. If you asked the people who use our service, they probably wouldn’t mention a lot of this. You don’t consciously evaluate good design, and it’s often hard to put into words. It just works really well.

Karen Hentschel: This is how we measure ourselves: do we get criticism, or do people say that what we do helps them?

Change, rules and outlook

Andreas Dahrendorf: How do you deal with constant change? At the moment, it feels as if a new model comes along every two weeks and you have to prompt it differently.

Marc Dantas Tiedemann: Fortunately, our software doesn’t change on its own, pick the latest models by itself or need constant adjustment. It’s set up in a fixed way. Of course, we keep an eye on the scores to make sure the quality doesn’t drift. But yes, things are changing all the time, new models are coming out and we evaluate them continuously too.

Another factor is that the models we use eventually get switched off. It’s crazy how quickly that happens. This really hasn’t been going on for long: our product has been live for a year and a half, and the first models we’re using are already being switched off, I think across several generations. Then they tell you: those are old now, you have to use the newer ones. So we really have to iterate again, even though we were happy with the output. That’s happening again right now. We have to keep adapting and find ways to make the evaluation processes faster and more automated.

I think we’re all in the same boat right now, whatever industry we’re in. So much is changing all at once. But that has a lot of positive effects for us too. For example, we’re fundraising at the moment, actively looking for investors. We need to prepare all sorts of documents for that, from financial plans and models to reports. What we’re now able to create with AI is fantastic. New things keep coming along very quickly that we can use, such as image editing for social media. So you can’t avoid having to keep up with things on every front.

Andreas Dahrendorf: A similar question, this time about the legislation and administration around it: where do you run into problems, what helps you, and what changes have you seen over time? Has the law caught up enough to respond to this area, or even take action?

Marc Dantas Tiedemann: Yes. For one thing, there’s the new AI law, the EU AI Act, as a framework. The regulation is now coming into force in stages and at least requires transparency. I don’t want to go into too much detail, but the main questions are whether you’re using AI in a high-risk area and how transparent you are with your users. We have to look carefully at what applies to us, especially since we may eventually be a medical device.

Then there’s the Medical Devices Regulation we’ve already mentioned. It doesn’t change very often, but there are updates from time to time. Witold keeps on top of it. He’s also responsible for medical devices at his clinic and knows the area very well. He monitors all of it for us so we can see what else we need to consider.

There are a lot of people who say: terrible regulations, Europe and Germany, everything’s so complicated, and then there’s the GDPR as well. Of course, we sometimes roll our eyes too when our lawyer tells us what wording to put in our tick boxes and how long it needs to be. It’s hard to expect anyone to read that, and whether it’s always sensible is another question.

On the other hand, there isn’t currently a comparable service in Europe that explains and supports patients, at least from what we’ve seen. There are already similar services in the US. One of our assumptions is that many people here don’t want to take the risk because regulation is stricter: you have to comply with the MDR, and you have to comply with the AI Act. As a start-up that understands this area, we have a lot of experience from our previous work, can navigate it well and have an advantage as a result.

We are also aiming for certification next year, with new products we’re currently building that will let us provide more guidance. That sort of thing can be certified. It also gives us a stamp, a seal of approval that sets us apart from the chatbots out there.

Karen Hentschel: On the other hand, you mentioned the ePA earlier: it would obviously be great if we could get to that data more quickly. It’s highly complex. From a purely innovative perspective, there are certainly lots of moments when you think: why can’t we access it more quickly? But from an individual’s perspective, it’s great that you’re in a protected space and that health data is handled the way it should be. That’s a very important asset in our system.

So, on one hand, as Marc said, we have a service that’s hugely relevant across Europe because it’s built on this quality standard. On the other hand, we need to get closer to the people driving innovation at a system level, so we can move solutions forward more quickly there. A lot is happening. There are great initiatives like SPRIND, which deliberately invests in breakthrough innovations and looks for authorities that are ready to say: we’re doing this now. That’s the mindset. We now also have a Federal Digital Service, which at least gives us hope that we can get closer to the British model and bring innovation in earlier and more deliberately.

That’s also why we do B2C: people need help now, and we can’t wait for these problems to be solved at a systemic level. So we’re doing it now. At best, that will influence things at other levels too. That’s usually how it goes: when innovation overwrites reality, it may eventually become the new reality.

Andreas Dahrendorf: And if you let your imagination run a little, maybe even dream a bit, where will you be in a year’s time?

Marc Dantas Tiedemann: Maybe we’ll be able to pay ourselves a salary by then. That would be a good start.

No, as I said, we’re actively fundraising right now. We didn’t want to for a long time, but now we’re looking for investors who are a good fit. We’re already having some exciting initial conversations, including with business angels, doctors who would invest, for example. Once that’s done, we’ll have more capital to develop faster. We have lots of plans and, of course, want to put them into action. We want to launch the app and reach more users. With our limited resources and just the four of us, everyone has a role, but we’re constrained in how much we can do overall. We want to move faster and reach more people.

Karen Hentschel: And the market is really exciting. You can see it everywhere, especially in prevention, where a lot is happening. It’s great to be part of it right now. I hope the landscape will have developed significantly further by next year.

My goal is for us to be able to support people more holistically with an app next year. At the moment, they have to come to us with this one document, and if they lose the translation, they have to get it translated again. I’d like us to support them over the longer term, do more for them, work more closely with them and create holistic value as a result, including around the system’s bigger questions. I think that would be great. And I think it’s achievable, given everything that’s happening right now. At the moment, there’s enormous potential to create something and make a difference.

Marc Dantas Tiedemann: Exactly. To build on that, one of the big goals we’re already working on is partnering with a health insurer. But the wheels turn slowly in that area. We see insurers as really strong partners because we have the same mission. One insurer once said: healthy patients are good patients. And our guiding principle is: health begins with understanding. Those ideas fit together very well.

Also, only doctors and insurers can access the ePA. With an insurer, we could take a much more targeted approach, for example, by connecting directly to the ePA in the patient portal and translating and explaining what’s in it. That’s interesting for the insurers too, because then the ePA can finally become useful. People will only look at it if they understand what’s in it. But it takes time. If we manage to do that by the end of next year, it would be a great achievement.

Andreas Dahrendorf: Have I forgotten anything? Is there a topic you’d still like to mention?

Karen Hentschel: I was thinking earlier: we’re here on EU AI Talk. Given that we’re working in this complex medical field, it’s worth saying that we could only exist in this form because AI now has these capabilities. That’s the only reason we could create the solution we have now with such a small team. In the past, what we’re doing now would have taken a hundred people with different tasks and responsibilities. I find it really exciting to consider what we can get out of these tools, despite all the criticism of the current moment. And then, looking at the market here, to ask how we can drive innovation forward more quickly. That would be great, of course.

Andreas Dahrendorf: Thank you very much for the conversation.

Karen Hentschel: Thank you. It was lovely to be here.

Marc Dantas Tiedemann: Thank you. I really enjoyed it.