My role and contribution
In our PhysioAI project I was responsible for the economics and the future-market test. The economics covers what we charge, which costs we have to cover and how many practices we need to break even. The future-market test asks what happens to our design once every patient, practice and competitor has access to capable AI agents. In the final deck that translated to three main slides and two appendix slides, all which had to fit with Jara’s problem and ecosystem work, Nadine’s prototype and Bence’s business model.
I ended up with this part after meeting up with one of my group members following one of the lectures. The assignment was split into nine components for four people, so everyone took one of the sessions plus another part. Since two of my group members had already chosen to start on their section I focused on session 4, mostly because it looked different from the rest and I found it interesting.
Important decisions and iterations
The decision that shaped my part most was how to offer and sell our product to practices. At the start we weren’t sure how to price or position, so I wrote down three possible models: practices pay per patient, insurers pay and patients get it free, or a staged route where practices pilot first and insurers pay once the value is proven. One tempting pitch we explored early on was that physios are in short supply, so a tool that frees up their time would let them take on new patients. Whilst the group found sources to support this claim, we worried it overstated the situation and did not want to build our whole argument on it. Most practices also charge per session so a tool that cuts appointments cuts their income, furthermore, blended care was not more cost-effective than face-to-face care in low back pain (Koppenaal et al., 2023). We ultimately agreed to stop promising fewer sessions and the value became physio time freed, about 12 minutes per active patient-month, which only counts if the practice can fill that time. This is why we chose the staged model and why we charge per active patient instead of per physio (the price follows usage and our cost to serve).
What surprised me
The break-even early version estimated we would need about 30 practices, which sounded very achievable. When I rebuilt it from the ground up (accounting for €13.50 in contribution per patient-month, 50 active patients per practice and a €400k fixed-cost plan), the scale grew to roughly 49 practices at €15 and about 78 at the €10 pilot price. That is around 0.5% to 0.75% of the nearly 10,400 practices in the Netherlands, a less flashy but much more defensible result. In addition, the break-even is also strongly dependent on the assumption that practices will monitor 50 patients. Lastly, the costs behind such as deployment costs are team estimates, so we show this range on the slides instead of defending one number. Nonetheless, some inputs do come from reliable sources like the €77 per hour physio rate witch matches the 2026 average for self-employed physiotherapists (Knab, 2026).
It is also easy to assume the AI is the expensive part when designing a strategy centered around it. This was another surprise as removing the language model cost completely would cut our running costs by only about 13% whereas supporting and onboarding practices costs more.
How feedback and evidence changed the work
Our own evidence changed my slides more than anything else. In our benchmark, a general-purpose AI without our grounding or computer vision agreed with the professional label in 2 of 4 mixed-label comparisons and never once said it was unsure. Today this remains insufficient as an alternative, but will certainly improve in years to come. So in my future-market slide I do not claim our AI model is free from risk, instead I propose to re-run the benchmark whenever a major model is released.
The same happened during the first draft of our project, where we still needed to come up with a pricing strategy and approach to the problem. After doing some individual research focused primarily on my parts, I shared my findings in the groupchat and proposed a B2B2C approach similar to that of the Tessa guest presentation. Jara answer with a different idea and after some fine-tuning we agreed on the approach demonstrated in our slide deck.
What I learned from collaboration
Dividing the work was the easy part, the hard part was that everything depended on other sections. My price depended on Nadine’s results and on how Jara described the problem, and small details started to contradict, repeat or blur across our work. For example, my pricing slide said video would stay on the phone and only measurements would be sent on (to stay within GDPR) while a teammate mentioned storing and interpreted it differently. Both made sense because the prototype uploads video for the demo while the intended product would analyse it on the phone, but it had to be made explicit. We met online a few times to agree how the deck should be structured, and again afterwards to check that it fit together. Going back through everything with a fine brush after we had all worked separately took longer than I expected, and I got into the habit of raising these things in the group chat whenever I spotted them. Next time I would agree a short list of shared facts and numbers at the start. During the presentations I felt comfortable with my slides but unfortunately the time was too short for all our parts plus a live demo of the prototype.
How the project changed my understanding of AI and business strategy
Before this project I thought of AI strategy as mostly a technology question, but it turns out to be more about who gains and who pays. Our own slides say it clearly, if appointments fall the practice loses revenue while insurers might capture the savings. So creating value and capturing value are different problems, and the staged model is our attempt to solve that. I also learned that the model is rarely the advantage since they are usually accesible and the main threat to us is practice software vendors bundling similar features. As such, any lasting advantage has to come from evidence, workflow and trust. Finally, I came to see how much of a business case depends on assumptions not tested yet. If I did it again, I would compare our pricing to a real practice before building the model.
References
Knab. (2026). Wat verdient een fysiotherapeut (zzp)? Bekijk uurtarief en winst in 2026. https://bieb.knab.nl/wat-verdient-een-fysiotherapeut-zzp-bekijk-uurtarief-en-winst
Koppenaal, T., van Dongen, J. M., Kloek, C. J. J., Arensman, R. M., Veenhof, C., Pisters, M. F., & Ostelo, R. W. J. G. (2023). Effectiveness and cost-effectiveness of a stratified blended physiotherapy intervention compared with face-to-face physiotherapy in patients with nonspecific low back pain: Cluster randomized controlled trial. Journal of Medical Internet Research, 25, e43034. https://doi.org/10.2196/43034
