Throughout this project, my main contributions were researching and validating the problem and opportunity, developing and evaluating our prototype, and analysing the competitive ecosystem and potential platform positioning. These responsibilities challenged my assumptions about AI and business strategy.
Decisions, Iterations and Learning
Our team started with a teammate’s personal experience of the information gap between physiotherapists and patients during home rehabilitation. I researched whether this problem existed more widely, both internationally and in the Netherlands. However, I realised I was primarily looking for evidence supporting an idea we had already selected. Recognising the risk of confirmation bias, I then expanded my research to investigate other industry challenges.
This revealed alternative opportunities, particularly around treatment costs and insurance coverage. Nonetheless, we retained our initial focus because the evidence supported the information gap and we realised the value of in-person physiotherapy. I questioned whether we had identified the strongest opportunity or simply validated our preferred one.
Having studied International Business, I already understood the importance of validating opportunities. What this experience made me reconsider was the distinction between validating a problem and comparing it with other problems worth solving. If I approached the project again, I would investigate competing industry bottlenecks before becoming committed to one solution.
Another important decision was taking responsibility for developing our working prototype. My teammate had created a mock-up, but I realised that an interface alone could not demonstrate whether our proposed movement analysis was technically feasible.
The mock-up suggested that AI could determine whether exercises were performed correctly or incorrectly. Through further research and development, I recognised that identifying differences between movements does not necessarily establish clinical correctness. We therefore shifted towards providing observable movement comparisons rather than unsupported clinical judgements.
Developing the prototype was challenging because I had no programming background. Although I had previously worked on digitalisation projects and knew AI could assist with programming, I had not realised how advanced its capabilities had become. By exploring AI development tools and learning to use GitHub and Render, I built a mechanism that compared a reference exercise video with a patient’s recording and generated movement feedback. After discussing the initial version with my teammates, I expanded it to demonstrate the patient and physiotherapist views and their feedback loop.
AI enabled me to create something I would not previously have been able to develop independently. However, I also realised that using AI did not remove my responsibility. I still needed to define what I wanted, evaluate the results and correct problems. This changed my understanding of how accessible technical development has become, while reinforcing the importance of human judgement.
A particularly valuable experience was designing the prototype evaluation. Although I had experience with primary and secondary research, I had never evaluated a prototype I had developed myself. I therefore researched how to make our testing meaningful and used exercise videos from a peer-reviewed study to compare three approaches: AI alone, AI with physiotherapy-specific grounding, and AI supported by MediaPipe.
We initially wanted to establish whether MediaPipe was necessary. If AI could already provide useful movement assessments, introducing additional technology might create unnecessary complexity.
The results surprised me. Among the four cases where label agreement could be evaluated, physiotherapy-grounded AI produced three label-consistent selections, compared with two for the MediaPipe-supported approach. MediaPipe provided more detailed movement measurements, but this did not improve agreement in our small benchmark.
The additional information was not necessarily useless. It could still help make assessments more transparent or support a physiotherapist’s judgement. However, our evaluation could not establish whether that information would improve clinical decisions. We also considered whether abstaining might sometimes be preferable to making an unsupported assessment, although our results could not determine whether the abstentions were appropriate.
This made me reconsider what a successful prototype evaluation should demonstrate. Initially, I was mainly interested in whether we could build the technology and assess movements. Afterwards, I understood that testing should also inform strategic decisions about which capabilities are worth investing in.
My main learning was that more detailed information or more advanced technology does not automatically create greater value. In future projects, I would first establish which decisions the technology needs to improve, then evaluate whether additional capabilities justify their costs and complexity.
AI and Business Strategy
Besides developing the prototype, I analysed PhysioAI’s ecosystem, competitors and potential platform positioning. Before this project, I regularly used digital platforms but had not fully considered how network effects and different revenue models influenced their strategic positioning.
We explored whether PhysioAI could become a platform by expanding into football and other sports clubs, for example by assessing movements during warm-ups. However, we realised that introducing additional user groups would not automatically create meaningful network effects. More importantly, this expansion moved us away from our original rehabilitation problem. We therefore decided not to pursue it.
What surprised me was how closely my ecosystem analysis connected to my teammate’s work on the economics of our solution. Through our discussions, I realised that positioning PhysioAI as an application rather than a platform had implications for how we could capture value, who would pay and how the solution would fit into existing physiotherapy practices. Even if patients benefited from better support during home rehabilitation, this would not automatically make the application attractive to physiotherapists. We also needed to consider whether it would create additional work and whether practices had sufficient incentives to adopt it.
This made the difference between creating and capturing value much clearer to me. I had previously understood these concepts separately, but working on PhysioAI showed me how closely they depend on one another. Our competitor analysis also made me realise that we should have investigated existing solutions earlier. Finding a real problem does not mean there is no existing solution, and differentiation needs to be considered before becoming too committed to developing an idea.
Collaboration, Feedback and What I Would Take Forward
One thing I learned from our collaboration was that trusting my own judgement and listening to others are not contradictory. Originally, I suggested using TikTok dance videos to test movement analysis, but my teammates questioned the complexity. After further research, I realised that it could distract from the specific movements we needed to assess and focused on repetitive physiotherapy exercises instead. Feedback also helped me recognise that my initial prototype needed to demonstrate the interaction between patients and physiotherapists more clearly.
Although we divided responsibilities relatively easily, our findings were closely connected. The overlap between my ecosystem analysis and my teammate’s economic evaluation showed me that sharing progress was not always enough. Discussing the implications of our findings was equally important. I think we collaborated well, but more frequent, in-depth discussions could have helped us recognise these connections earlier.
Thus, the AI Strategy Lab changed how I would approach future AI projects. I would investigate alternative problems and competitors earlier, consider adoption and business-model implications alongside development, and use evaluation results to challenge technological investment decisions. Finally, my biggest takeaway is that developing an AI strategy is not simply about identifying what technology makes possible. It is about questioning whether the solution addresses the right problem, creates meaningful value and justifies the resources required to develop it
