My role and contribution
In our team project we developed Crit Happens, an app where an AI acts as the Dungeon Master for groups that want to play Dungeons & Dragons but have nobody to run the game. My part was mostly the economics. I worked out what it would cost to set up and run the product: so the server costs, development costs and monthly maintenance, and I made the cost slide for that. I also took the lead on the subscription model. I decided which tiers we would have and what would be included in each one. Based on that I built the monthly scenarios with subscription revenue, ad revenue, operational costs and the expected profit. Later I also made the future-market test slide. Next to my own slides I helped a lot with the layout of the complete presentation.
Before all of this, we of course sat together to come up with an idea. Our first idea was about stock management for small supermarkets, so they could prepare better for changes in what customers need. After the first week we discussed it with the lecturer and it became clear it was not a strong idea. So we switched completely. A teammate came up with the AI Dungeon Master and the whole team helped to shape it. Looking back, that switch was one of the best decisions we made, because the project became a lot more fun to work on and I think that made our work better, too.
Decisions and iterations
The biggest moment for me was when I first ran the numbers. The low and medium scenarios were making a loss, and at first, it was even worse than what is in the final slides. That was a bit of a shock. We discussed it as a team and looked at two options: raising the prices or adjusting the expected number of subscribers. We decided against higher prices, because we wanted the product to be attractive for as many people as possible and we were afraid to lose customers. Instead, we raised the expected number of users, because our first estimates were very conservative for a product that can be used worldwide. Looking back, this was the easier solution and maybe not really a fix. The low and medium scenarios still lose money, and the profit in the high scenario depends on growth we have not tested.
The server choice also changed during the project. At first, we were convinced we needed our own server. With an API like Claude or ChatGPT the costs go up with every user, we would be limited in what we can do with the model, and switching later gets harder the bigger you get. I worked out two options: renting a server, with low upfront costs but costs that keep growing with usage, or buying one, with around €50k upfront but much lower marginal costs per user. We first went for buying, because we expected to scale a lot. But next to the low and medium scenarios, €50k upfront for a product without proven demand is a big risk. In the end we recommended to start on the Claude API, so we have no upfront costs and can test the market first, and to switch to our own server once demand is steady. That switch will probably bring its own problems, but with what AI can do now I think it is manageable, and it is better than paying for a server nobody uses yet.
What failed and how feedback changed the work
We did not get many formal feedback moments during the course, so most feedback came from the lecturer at the start and from each other. Some of it was about my own mistakes. A few of the sums in my economic slides were wrong. A teammate noticed it and we corrected them. A bigger mistake was the future-market test. I had missed that this slide had to be about AI agents, and I spent a lot of time on a completely different version that turned out to be useless. That was really my own fault, I should have read the assignment better before I started. After that, we rebuilt the slide around what happens when players, Hasbro and the platforms all get capable AI agents, and it is a much stronger slide now. At the end we also asked a few LLMs to check the deck for inconsistencies between the slides. They found some, and we fixed those ourselves. We did not let AI write big parts, because we wanted it to stay our own project.
Collaboration
The collaboration went well in the beginning. We had meetings, we knew who did what and we stayed in contact through the group chat. Near the end it got harder. I was a bit late myself because work got in the way, but I told the team in advance. One teammate did not do that. He had nothing ready at our internal deadline, worked on his slides on the day itself and went offline two hours before the real deadline while there was still a lot to do. The other two teammates and I finished the slides together. It was frustrating, but it also made me look at myself, because I have been on the other side of this before and I am not proud of that. Since then, I try much harder to stick to what we agree on as a team. Next time I would take the internal deadline more seriously, and if something is not done, ask directly why and what that person needs, instead of accepting a simple “I’m going to do it.”
How my view on AI and business changed
Before this project I mostly saw AI as a nice gimmick for businesses. Now I see that you can build a whole product with AI in a way that was not possible before, and I expect a lot of software will be made like this from now on. At the same time, the economics surprised me. AI is a lot more expensive than I thought for startups, and the upfront costs are a real risk. Ads also brought in much less than I expected. And because competitors can use the same models, the AI itself is no moat and you cannot ask high prices. That is why our slides say the edge has to come from play tested stories and a rules engine the AI cannot overrule. I did not think about it that way at the start. Later on, data from real sessions could help to improve the stories, but only with consent and the safeguards from our governance slide. I also notice that a lot of software made with AI starts to look the same.
AI still makes mistakes, people are not always open to it, and someone has to be responsible when it goes wrong and steer the product in the right direction. AI makes building faster and cheaper, but deciding what to build, what it may cost and who carries the risk is still up to us.
