Project Learning Blog – Milan Girbaran

No ratings yet. Over the past weeks, my group members and I have been building Shiftable, an AI staffing coordinator for Dutch restaurants and hotels. Someone calls in sick, the agent finds a replacement and confirms it, and the manager only has to deal with the stuff the agent can’t sort out. My part was the competitive and ecosystem analysis, so who we compete with, what we depend on, where we sit as a platform and what we could actually defend. Max built the prototype and did the business model together with Karmen, Karmen also did the future market test, and Joris did the first drafts of the economics and the evaluation.
The project didn’t start with my idea. We all pitched one and mine was an agentic lost and found service, but we ended up going with Karmen’s hospitality idea. Our argument was that lost and found was too broad and hard to test, while a hospitality agent you can just simulate. Fair enough honestly. Looking back, that was a good lesson: a cool idea isn’t worth much if you can’t really test it.
What we ended up handing in looks pretty different from our initial idea. In the proposal we wanted a RAG onboarding assistant that new staff could ask stuff like “how do I void a transaction?”, and our key assumption was that managers lose a lot of time answering the same questions over and over. My first competitive analysis was written for that idea and the first thing I found was that it already exists. Connecteam already sells an AI agent, an AI course builder and auto scheduling in one app, and Oneteam (a Rotterdam based company) already has an AI chat over a venue’s own documents. So we weren’t creating anything new, we were walking into a pretty crowded market.
After a while we decided to narrow down to autonomous staff scheduling, with way less manual work than Shiftbase. I agreed pretty much straight away, which also meant most of my analysis had to be redone. We never actually tested the assumption from our proposal, we just dropped it when the idea changed. It was probably the right call, but it does show how easy it is as a team to pivot away from a question instead of answering it.
The pivot also raised the question of whether we were a plug-in on top of Shiftbase’s API or our own platform. In the end it was kind of both: a standalone app for venues still running on WhatsApp and spreadsheets, and an agent layer for venues that want to keep Shiftbase, which we framed as an on-ramp to the standalone app. I started on the competitive part before the business model was really settled and that cost me a few rounds of rework. You can’t really do a sharp competitive analysis before you know what you’re selling.
The research did change the argument quite a bit. Once I found out Planday launched AI shift assignment in March 2026 and Shiftbase already had forecasting live and intelligent scheduling on the way, it was pretty clear “AI scheduling” couldn’t be our moat. So the question became what does our product build up that someone else can’t just copy. At first I had a Dutch rules engine down as an advantage, but that doesn’t really hold up against Shiftbase, which is Dutch and knows the horeca cao at least as well as we would. In the final deck we called it temporary. What we ended up calling our strongest edge is the venue data the agent learns from, so which staff members take which shifts, on which days, with how much notice. A competitor can copy the feature in a few months but not that history, and if a venue leaves they start from zero. This way we make use of vendor lock-in.
Working in a team of four also showed me how much the parts depend on each other. A slide can be right on its own and still clash with the rest of the deck. For a while our slides didn’t agree on whether we were a platform or an integration layer, the name switched between ShiftWise and Shiftable, and the problem slide mixed US and Dutch numbers. We fixed it all on the final evening, but next time I’d read the other slides as part of my own work much earlier.
The biggest thing that I learned is how I think about AI and strategy. At the start I kind of assumed the AI was what made our product special, and now I see it more or less the other way around. Every competitor can use the same models. The value is in what’s around the model: data that only builds up by using it, integrations that make it annoying to leave, and the trust a manager needs before letting software act for them. The prototype showed the same thing from another angle. Max found that the system works best when the model is only used at decision points, like choosing who to ask, and everything else just runs on fixed rules to save usage tokens. That kept the costs low and gave the AI fewer chances to mess up. So using less AI actually made the product better.