Student Role and Contribution
In my team with Maximilian, Joris, and Milan, my main responsibilities were the business model and the future-market test. For the business model, I worked through how Shiftable creates, delivers and captures value. For the future-market test, I examined how our strategy holds up if customers, competitors, platforms and staff all use capable AI agents. Although I mainly worked on these parts, we all reviewed each other’s sections throughout, and our final solution was a collaborative effort.
Important Decisions and Iterations
The biggest iteration came early. After we each came up with ideas, we went with my idea focused on the hospitality industry. My first proposal had two agents: an onboarding chatbot-style agent for new staff and a scheduling agent for managers. I had thought about how exhausting onboarding is for hospitality workers in high season, and how scheduling is a headache when many employees work part time. It didn’t take long for the idea to change.
After conversations with people we knew in the industry, we discovered that businesses document their policies very differently, and the value of an onboarding agent was hard to measure. A manager might spend more time updating the agent than answering questions directly. We briefly considered AR/VR for onboarding but dropped it as too complex to standardize. In the end, we focused on scheduling for Dutch restaurants and hotels with up to 40 staff. Unlike bigger businesses with more sophisticated systems, mid-sized businesses often schedule in spreadsheets and group chats. They already navigate simple digital tools, so this became our key target market.
These conversations also helped us iterate our design. People described last-minute notice, slow manager replies and the stress of calling in sick. When we proposed our idea, some said it is likely easier to turn down a shift when an agent asks than when a manager does. This made life simpler for employees but could complicate things for a manager who needed someone to work. So we built an escalation rule: the manager is notified when half of the eligible staff decline, or when nobody accepts two hours before the shift. These discussions also helped us simulate staff behaviour.
Surprises throughout the project
What surprised me most was how small the improvement in shift coverage was. In our simulation of 1,000 absences, the agent covered 58% without the manager. Shifts covered before they start, including manager follow-up, reached 85%, against 82% for a group chat. I had expected the benefit to appear as more shifts being filled. Instead, the value came from manager time: hands-on time per absence fell from 10 minutes to 3.5. Pitching our value as filling more shifts would not have been supported by our own evidence.
Another surprise was the ROI on our model. Two teammates found different statistics for a manager’s hourly wage. After checking the source, we decided on the higher figure of €31 an hour. Even so, if the agent saves only half the time we expect, venues get back at least €1.70 for every €1 they pay.
How Feedback and Evidence Changed the Work
For my future-market work, I tested platform envelopment (Eisenmann, Parker & Van Alstyne) against our business model. From course readings, I knew a larger platform can absorb a smaller product’s function. This matters for us because Shiftbase is both an integration partner and a competitor. My team challenged this because we have no engine to license. The language model and the agent software are interchangeable, and what a rival cannot easily copy is each venue’s own data, such as who accepts which shifts on which days. After discussing it with the team, the licensing recommendation came off the slide. Instead, we focused our competitive edge on venue data, though I kept envelopment as a named risk.
What Was Learned from Collaboration
We had a nice mix of academic backgrounds. Mine is in business and psychology, so I was comfortable with the business model, while Maximilian, who did a minor in business information, was strongest on the prototype and its running costs. We divided sections, but the most valuable moments came when we met, since each component was intertwined. My price depended on time saved, which depended on how often the agent handles an absence alone. Brainstorming together and comparing findings showed me how the parts fit.
I knew very little about software and AI, so my learning curve was steep. Constantly debating ideas with teammates who I believe had a stronger AI background taught me a lot, such as how the costs of different AI models like ChatGPT and Mistral affect a business’s bottom line. We struggled with planning at the beginning, but flexibility and open communication got us to something we were proud of.
How the Project Changed My Understanding of AI and Business Strategy
Before this project, I thought of AI as technology that speeds up a workflow from end to end. That thinking was challenged, and I realized AI implementation is not that simple. A business can easily implement AI, but that doesn’t necessarily make it more productive or worthwhile. Sometimes implementing AI in a single part of the workflow is all that’s needed. Challenging our own project forced us to think about why a business would want our model specifically.
I also became more aware of human costs. On paper, the savings looked hard to beat. In practice, a manager may need more mental energy to check a generated schedule than to build one. Staff at small venues might feel like a family and may not want their schedule from an AI system, which affects acceptance.
The strategic question I ended this project with is this: if the tasks AI automates most easily are also the easiest to copy, and automation moves effort onto people, where does lasting advantage come from? My work on the future-market test helped me understand this question. AI coordination itself is not a competitive advantage, since platforms are likely to add agents at no extra price. The stronger position comes from accumulated venue data, feedback and integration into daily operations, which raise switching costs the longer a venue uses the product.
Looking back, this project taught me to ask where AI belongs before asking how to build it. Our strongest results came from using it sparingly, measuring its value in manager time, and treating feedback and evidence as reasons to change course. If I were to repeat it and if we had more time, I would test the human costs earlier, since those are the assumptions our simulation could not settle.
