Starting out with our project, I took the role of developing the working prototype. Our role distribution, however, was not specifically tied to one group member, and that was on purpose. We worked through every requirement of the project together, as we thought the opinion and input from every member would help create a more unified vision for the project. Nearing the end of the project timeline, the role responsibilities materialized, as each member shifted towards the part they seemed more proficient in.
The Initial Idea
Shiftable at first had a very different problem to solve. Our initial idea was to solve the information gap that exists during staff onboarding. My group member Joris, thanks to his experience in hospitality, outlined that an AI-powered knowledge base could help with onboarding and answering questions for new staff members, helping relieve more experienced staff and making onboarding a more pleasant experience inside hotels and restaurants. That’s when we also mentioned a staff scheduling tool, though only as an add-on.
The Pivot
After conducting interviews with people directly involved in the hospitality industry, one thing became clear to me. The questions we imagined new staff faced had less to do with the processes in the industry and more to do with the tools and programs used on the job, such as hotel- and restaurant-specific ERP/CRM systems. While creating an AI-assisted knowledge base would definitely add value to said industry, the cost would simply not make it seem valuable enough for them to implement, coupled with the fact that those software providers could also launch their own AI helpers, which we couldn’t compete with since those providers know their systems far better than we ever could. This was exactly when we decided to focus on the other pain point of the hospitality industry, which was staff planning. That, however, was also not without challenges. At first we thought the idea was dead on arrival, as there were big players already on the market, such as Shiftbase. After careful consideration by the whole team, we did find a niche Shiftbase was not yet covering: fully AI-assisted scheduling, where an AI agent handles everything from staff schedule creation to reaching out to employees to fill gaps.
Building the Prototype and Issues Encountered
Building the idea seemed relatively straightforward. The manager would set the employee information once on our dashboard and the AI would take it from there. However, when I was calculating the cost breakdown, having an AI agent handle the whole process chain quickly became economically unviable. LLM API calls quickly ate up our already slim margins, coupled with ever-rising server infrastructure costs. My initial idea was an option for the manager to simply connect their ChatGPT account to the agent, thus shifting the AI cost to the venue itself. But this idea was quickly buried, since this would introduce serious GDPR concerns, as we would be sending confidential employee data to servers outside the EU. Thus we settled on using Mistral’s AI model, which was more expensive than US models, yet fit our GDPR and AI capability requirements perfectly. To minimize API costs as much as possible, most of Shiftable works purely on the program logic alone, using AI only at “decision crossroads”, where an LLM’s reasoning is needed, such as interpreting free text from employees or edge cases such as moments where no suitable employees are available and the agent has to decide what next steps to take, of course inside a ruleset I created. This doesn’t make Shiftable any less of an AI agent, though: employees can still message it through any channel and it can interpret those messages and act on them through its tools, namely the schedule and the messaging channels. The program logic simply takes the heavy lifting off the LLM, for example by pre-selecting which employees actually fit an open shift through our ruleset so that the agent only reaches out to them instead of messaging every single employee. Because data security and privacy were part of the architecture from the start, we did not face significant issues later down the line; however, the constraints from those decisions did affect the predicted profitability of Shiftable.
Testing and Feedback
While my teammates focused mostly on running simulations to calculate the actual time saved on scheduling through the AI agent, I took a more usability-focused direction, asking actual managers in the industry for their opinion on the product. One major piece of feedback stood out the most: while many did see the benefit of Shiftable, some did not fully trust the agent’s decision-making. Because of this, I added options to hand more control to the managers, should they need it, like manually approving AI-generated schedules or reaching out to employees directly through the dashboard. However, we deliberately did not give managers too much control, so as not to hand operations over to them completely, as it would eliminate the time and cost savings introduced by our AI agent, since our value was mainly calculated through time and money saved on the staff handling/schedule creation process.
How My Views Changed
Due to prior experiences with similar projects, I approached this assignment with the mindset that not everything “needs” AI. Working on this prototype reinforced this assumption, since it made the whole team go from thinking where we could use AI to where we actually needed to. With many people already distrusting LLMs due to their potential unreliability and trusting AI agents even less, a targeted and directed use of AI is needed to reduce issues such as hallucinations. Later down the line, we realized that our long-term edge was the data we would collect from the venues rather than Shiftable’s architecture, as the LLM (in our case Mistral) or the agent harness (Hermes) could both be swapped, but the venue-specific data is what would remain. The AI agent could synthesize patterns such as employee schedule preferences, common gaps in the schedules and when they would occur, helping staffing managers spot mistakes they would otherwise miss, as they had no way of accumulating that data. This became our way of introducing self-reinforcing feedback loops.
Finally, I’d like to think that Shiftable’s value proposition is obvious to employees, staff managers and owners, but our biggest challenge still awaits us: how willing people will be to accept and follow instructions from an AI agent rather than a real person.
