My role
During the project, I had several roles. The first was taking the initiative on certain tasks, such as setting up most of our meetings, checking that we were all on track, and making sure we agreed on a plan before anyone disappeared to work on their own slides. When we didn’t agree, we talked it through until we did, and those conversations were usually constructive rather than awkward. To be fair, everyone communicated well, and I certainly was not the only one who took the initiative. The second role was more content-related, although I would argue this was not necessarily a personal role. Although, we made sure each slide/subject had an owner, we cooperated closely and reworked all of the subjects together, which means the end result was really a cohesive and ‘collaborative’ piece.
Pivot or stay?
The biggest decision came about halfway through, when we started doubting whether our idea was still relevant. Planday and Shiftbase were already announcing AI scheduling features, and for a while it felt like we were building something that our ‘competitors’ would ship next quarter. We seriously discussed dropping it and starting over with a completely different idea.
We decided to stay, but not unchanged. Instead of competing on “AI that makes schedules”, we focused on what actually set us apart: an agent that runs the whole coverage loop, from a sick call to a confirmed replacement, with the manager only stepping in for exceptions. We also leaned into the parts that competitors would find hard to copy. In hindsight this was the right call. Pivoting that late would have given us a shallow version of a new idea, and the doubt itself fed straight into our ecosystem slide.
Surprises and failures
For the evaluation we wanted something better than “our demo looks nice”. We couldn’t get a real restaurant to try Shiftable in time, so with help from an AI assistant that Max created, we ran our prototype’s actual eligibility and ranking code against 1,000 simulated sick calls and compared it with a manager coordinating through a group chat. The code was real, the staff replying to offers were not.
That test still caught something real. The prototype covered only about a third of sick calls without the manager, because it offered a shift to one person at a time and escalated as soon as someone stayed silent for 30 minutes. Our own design was the bottleneck, not the AI. We fixed it by offering shifts to all eligible staff at once, and in the re-run the agent covered 58% of absences on its own.
What ‘failed’ was getting out of the simulation. Every reply and acceptance rate in the model is an assumption, so that 58% says more about our design than about real hospitality staff. It’s the biggest limitation in our whole deck.
The second surprise was the cost model. My first rough estimate said AI was the cheap part and human support was the real cost. Once we rebuilt the model around Mistral and actual server prices – thanks to the great input of Max – inference became the biggest single line, about €18 of the €40 per month for a 10-person venue, growing with every employee. That flipped our thinking. Using the model only at decision points, and pricing in tiers by team size, came straight out of that correction.
How feedback and evidence changed the work
Two comments from teammates shaped the economics more than anything I read. One pointed out that a subscription only makes sense to a manager if you show it against their own hourly wage and the time it saves. That became our headline: every tier pays for itself if it saves a manager less than an hour a week. The other reminded me that the brief explicitly asks for human and AI costs, not just token prices. That pushed us to show onboarding, support, compliance and manager oversight next to the inference bill.
Checking our own evidence was humbling too. I noticed that the 8.4 hours slide was a US figure sitting under a “Dutch hospitality” headline. Although it is likely to be generalisable, we softened the headline in order to ensure consistency and reliability. We also questioned our €31 manager wage, because wages varied largely across different sources. However, the most reliable ones all stated wages around this price, so we kept it for consistency, as the conclusions hold either way.
What I learned
The main thing I took away is that agreeing on a plan early is much cheaper than fixing things later. The pivot discussion went well because we argued about evidence and positioning, not about whose idea was better. I think nobody walked away feeling overruled or not heard.
Where we struggled sometimes was consistency and planning . Splitting the deck by slides meant numbers and even the product name drifted between sections: we went through ShiftWise, ShiftAble and Shiftable at various points. At points, we did not have a set schedule of how to proceed, but on the other hand this also caused us to be flexible. Since we held regular meetings, this was not a big deal in the end as we managed to work on everything properly.
How my view of AI and strategy changed
I started this project thinking the AI was the product. I now think it is the least defensible part of it. Barney and Reeves (2024) make the point that AI amplifies an existing advantage rather than creating one, and our own work kept proving it. Any owner can open a general chatbot, and the platforms will bundle agents in for free. What is left is levels 2 and 3: our own rules, each venue’s data, and a loop that learns which offers people actually accept (Cook et al., 2024).
I also learned that value comes from redesigning the workflow, not from adding intelligence to it. Our best decisions were about where not to use the model: the rules decide who may work, the AI only steps in at real decision points, and the manager approves by default. That design was cheaper, safer and easier to defend under the AI Act, all at once.
Finally, I now treat strategy as something you test. A simulation is still only a simulation though, which is why our recommendation ends with a paid pilot rather than a launch.
References
Barney, J. B., & Reeves, M. (2024). AI won’t give you a new sustainable advantage. Harvard Business Review, 102(5), 72–79. https://hbr.org/2024/09/ai-wont-give-you-a-new-sustainable-advantage
Cook, S., Hagiu, A., & Wright, J. (2024). Turn generative AI from an existential threat into a competitive advantage. Harvard Business Review, 102(1), 118–125. https://hbr.org/2024/01/turn-generative-ai-from-an-existential-threat-into-a-competitive-advantage
