“Triathlons are never won during the swim, but they can be lost.”
This is a statement one of the team members liked to quote, and it became a guiding principle for our project. He has been a lifelong swimmer, so the team understood what good technique is worth and how often it is lacking, especially for triathletes. However, receiving professional feedback from a coach quickly costs over €100 per individual session, making it unaffordable for the majority of aspiring athletes. Considering the increasing popularity of triathlons and Ironman races, we decided to create a solution as our project.
Our solution is called SWYM, an AI agent that helps athletes improve their swimming technique. To make it work, the user records two videos from different perspectives, which are then uploaded to our agent. The agent first models the movement through pose estimation and then analyzes it. Based on a comparison with training data from professional swimmers and factors such as body size and proportions, it produces personalized feedback: which mistakes the athlete made and drills that can be practiced in order to fix them. The goal is to raise the user’s performance and help them avoid injuries at a fraction of the cost of traditional coaches.
Financials and Pricing
Within the team I was responsible for the financials: the setup costs, what one AI analysis costs, how much we should charge for it, and how profitable this venture could be overall. When researching the cost side, I initially assumed a fairly high analysis cost. The actual cost came down to roughly 8 cents per analyzed video, and only about 3.5 cents of that is attributed to the analysis itself, with the rest being used for the one-year storage of the videos. Furthermore, the pose estimation model we used in the prototype is even free to use for open-source projects, such as a university project, although commercial use requires a paid license. Therefore, the running cost of the AI agent itself is surprisingly low. This result became a crucial input for SWYM’s pricing model.
The most interesting part to me was the development of the pricing. Drawing on the readings of the class, I aimed to use uncommon but effective pricing techniques for AI agents. Since our goal is to improve swimming performance, I had initially looked into an outcome-based pricing model, based upon the improvement of technique after usage of the AI agent (Dharani et al., 2025). However, this idea came with a notable number of risks: How do we prove that the improvement was caused by SWYM and not by the extra training the athlete did? Does the client pay first and get reimbursed if it does not work, or only pay afterwards? How specifically is performance measured? Each of these questions could lead to a potential dispute with a user, and the uncertainty made financial estimations difficult. Therefore, I moved the pricing to a usage-based model. SWYM charges per interaction, with a monthly limit that depends on the chosen plan.
I designed three plans: a Free plan with a single monthly interaction to let users experience the product, a Pro plan with four monthly analyses to support regular training, and an unlimited Max plan for those athletes who aspire to train the most. Since athletes are likely to value feedback very differently, the three plans allow users to pick the version that best matches their needs and willingness to pay, making use of value heterogeneity. The Free plan is deliberately limited to one swim a month and only shows the biggest fault of the session, so that a serious athlete is incentivized not to stay on it. Creating three plans furthermore makes use of the compromise effect, where customers tend to pick the middle option when presented with three choices (Hanke, 2024). On top of that, I added a referral system, because athletes who refer each other are likely to train together and motivate each other’s usage, an idea inspired by Belo and Li (2022).
The biggest lesson in all of this was that the price should not follow the cost. Since an analysis costs 8 cents and on the Pro plan an athlete pays about €2.50 per swim, the pricing is not based upon the “cost per unit”. Rather, it is derived from what a beginner or an ambitious triathlete is willing to pay. More importantly, users are likely to compare the price of SWYM to the next best alternative, which is professional coaching that costs over €100 per single session. Overall, the low cost of the AI did not make pricing any easier, but it is a great example of why traditional cost-plus pricing fails for AI products (Dharani et al., 2025).
Persisting challenges
Even though the business case looks strong, there are still two critical challenges:
- The financial model estimates break-even at about 3,000 paying athletes, which looks very achievable. However, I had initially forgotten to include conversion and churn. Freemium apps typically convert only two to five percent of their free users into paying ones (Kumar, 2014). So, assuming five percent, those 3,000 paying athletes require around 60,000 free sign-ups. And since most users are unlikely to stay forever, a monthly churn of five percent would mean SWYM needs roughly 1,800 new paying athletes, or about 36,000 new free sign-ups, every year (Mobile App churn rate Analysis, n.d.). Therefore, the real challenge of the venture is user acquisition and, more importantly, retention.
- Every analysis needs a video of the user, filmed in a swimming pool. Most pools are public, so other swimmers might end up in the frame without agreeing to it, and many pools restrict filming for that reason. This could be solved by automatically blurring any other person in the video. However, pools might still forbid filming, making SWYM unusable in those cases.
Conclusion
The biggest lesson this AI project has taught me is how feasible it has become to build a product with AI at its core, and how far the technology has developed in recent years. Working on the financials also taught me something less obvious about strategy. Since the analysis has low costs and the tools for the model are technically available to everyone, the AI itself is not what makes SWYM valuable. The value sits in the feedback that was not affordable before, and in the AI agent’s learning curve, which improves the feedback with each analyzed video. Having done the financial analysis, I believe SWYM could make it to market. The AI cost is low, but the value provided is not.
References
Belo, R., & Li, T. (2022). Social referral programs for freemium platforms. Management Science, 68(12), 8933–8962. https://doi.org/10.1287/mnsc.2022.4301
Dharani, N., Pineda, J., Andeyro, A., Hamming, N., Kim, Y., Misra, N., & Bhatia, A. (2025, August 13). Rethinking B2B Software Pricing in the Agentic AI Era. BCG Global. https://www.bcg.com/publications/2025/rethinking-b2b-software-pricing-in-the-era-of-ai
Hanke, P. (2024, June 13). Compromise Effect. Buynomics. https://www.buynomics.com/glossary-blog/middleoption
Kumar, V. (2014, May 1). Making “Freemium” Work. Harvard Business Review. https://hbr.org/2014/05/making-freemium-work
Mobile App churn rate Analysis. (n.d.). Tenjin. Retrieved October 11, 2026, from https://tenjin.com/glossary/churn-rate/
