In this course our team made School IRL. It is an app that students can use their phone camera to scan everyday objects, and then the app connects the object to what they learn in class, with quests given by the teacher. In our team I was responsible for competitive and ecosystem analysis and the final recommendation. When I look back now, I think my biggest contribution was not adding new features to the product. It was more about finding reasons why our product could fail, and thinking about what we need to change.
Actually our first proposal was rejected by the professor. We wrote a lot about the features, like object recognition and the teacher dashboard, but we did not really explain what problem we want to solve and why we need AI for it. To be honest, at first I thought the idea is obviously useful, so I did not understand why the framing is so important. Later I realized that “obviously useful” was just our assumption and we never tested it.
This feedback changed the way I did my part. When I start the competitor analysis, I did not only look for apps that look similar to ours. I tried to ask a different question, which is who is already doing the job that schools would pay us for. The result was not very good news for our team.
I found that recognizing an object and giving an explanation is not special anymore. Google already offers Gemini in Google Classroom for free, and since August 2026 secondary school students can also use it. Google Lens can also recognize objects. This means the main function of our prototype is something a much bigger company already gives for free, in a tool that teachers use every day. In the course we discussed about platform envelopment, and I think this is a clear example. Google can easily put a simple version of our idea into a system that schools already depend on.
So in our own presentation I had to say that our core technology is not our advantage. I thought what can still be hard to copy are three things: an object library that matches the Dutch curriculum and is checked by teachers, a way to prove the student really went to the place, and the teacher’s routine when quests become part of lessons and grades. When I tested this idea again, I found that the library alone is not strong enough, because once common objects are all mapped it will stop improving, and Google can also do it. That is why we suggested to build School IRL as an add-on for Google Classroom, not to compete with it directly.
Some evidences I found also changed our plan. In our proposal the key assumption was whether schools will accept students to use devices for these activities. When I mapped our dependencies, I found that from January 2024 Dutch secondary schools do not allow phones in classroom, except for educational use. This did not kill our idea, but we had to change the design. The quests should happen outside the class as fieldwork, and phones are only used in class when the teacher leads a discussion. Another finding was from the EU AI Act. AI that evaluates learning outcomes is seen as high-risk, but in our economics part we promised automatic grading. So we changed it, the AI only suggests a score and the teacher need to confirm it. For me the most surprising thing is that these two findings were not about the product itself, but they changed the product more than all our discussions about features.
For the final recommendation we decided to revise and run a pilot first, not to scale immediately. The pilot will test one topic in two or three Dutch schools for one semester. Classes will be randomly assigned to use School IRL or normal homework, and we wrote down the conditions for scaling before collecting any data. The most difficult part for me was the “stop” column. We decided to stop free scanning outside quests, gallery uploads and fully automatic grading. It was not easy, but it made us say clearly what our product is not.
I also made some mistakes. In my first draft of the scaling plan, I wrote that we will add biology and engineering quests and extend to grades 10 to 12. Both were wrong, beacuse our prototype already had biology and engineering quests, and our team never decided a target grade. I used an AI tool to help me draft these points and I did not check them carefully with our own materials. After that I also thought about expanding to professional training, but I realized it is a totally different customer and curriculum, almost a different business, and our pilot cannot prove anything about it. In the end we chose three directions: selling to school boards and municipalities instead of single schools, adding more school subjects with museums and local sites as partners, and letting teachers share quests between schools. The last one is also important because in my analysis the network effect mostly stayed inside one school.
Our team had four people and everyone was in charge of different sections. This made us work faster, but when we put everything together, some parts did not match. For example, different sections had different target groups, and one part said School IRL can replace field trips while my recommendation saw it as an everyday supplement. From this I learned that when a project is divided like this, checking if all parts tell the same story is also very important work. My defensibility analysis also overlapped with my teammate’s future market test, because we both thought the advantage should move to verification and teacher workflow. We agreed that I focus on the current competition, and my teammate extends it to the situation with AI agents.
About AI itself, our prototype was mostly built with AI tools and it really worked, which really surprised me. But when building become this cheap, the prototype is not what makes us different anymore. This is actually the same conclusion as my analysis about object recognition, but this time I experienced it by myself. Using AI for research also taught me a similar lesson. It helped me find sources much faster, but sometimes it wrote things that sounded correct but did not fit our project. The solution is the same rule we put in our product: AI can suggest, but a person who knows the context should make the final decision. Before this project I thought AI strategy is mainly about what AI can do. Now I think more about what is still scarce when everyone has the same AI, and who should be responsible for the result.
