Did you ever ask yourselves: “How can we translate the chaos of monitoring over 1,000 SKUs and unpredictable live variables—such as D1 highway closures into a transparent, actionable interface?”. That’s what we did! We tried to answer this question based on a gas station company operating on highways in Czechia – “MOL”. Let’s recap what we thought any gas station especially MOL situations are currently facing:
- The current System: A static rule someone at MOL created that states no order can exceed 12,500 Kč (€500) without the approval of a manager. You arrive at your shift as a manger of one of MOL’s gas stations (same as our target persona Marek Dvořák) to find empty coffee shelves because the system panicked at the price tag and did not proceed with the order. Some might say typically for a Monday morning.
When initially approaching the prototype, I felt overwhelmed by the sheer complexity of monitoring the comprehensive dashboard in real time. Accounting for dynamic updates—such as D1 highway closures or severe weather conditions resulting in delayed shipments—initially seemed a daunting challenge. So the question was unsolvable! Therefore, one of the challenges I faced were that there was no solution in sight. Nevertheless, my team and I deconstructed the challenge into distinct operational layers. We first analyzed the overarching strategic value for the MOL Group, then evaluated the direct practical benefits for the station managers, and finally determined how to effectively monitor the comprehensive inventory of a single site encompassing assumingly over 1,000 Stock Keeping Units (SKUs).
Fall seven times, stand up eight.
The first prototype I created – some might say it is a “Black-Box AI”. My, initial prototype design reflected this flaw: I built the logic so that the system would autonomously handle everything under the legacy 12,500 Kč (€500) limit, and only surface orders for manual review if it detected a validated external disruption (like a Modřice roastery breakdown – not able to serve our gas station with coffee beans as the heating coil just broke down).
However, this assumption failed upon contact with reality pretty soon. I was surprised to realize that by trying to build a “smart” system, I had accidentally created a rigid one. If the AI encountered an unprecedented order size with no contextual data, my initial logic would have let it slip through or block it without nuance. I learned that in real-world supply chains, the absence of data is just as critical as the presence of it. Therefore, let’s phrase it as it is a complete failure. No added value for MOL. Luckily, we did not give up at this point.
The Power of Collaborative Pushback
The turning point of this project was a direct intervention from my collaborators. During a review session, Michal (one of my team members) pointed out a massive blind spot in my UI copy and logic. I had explicitly written that only deliveries with validated supplier delays or highway detours would surface for approval. Michal sharply corrected this, noting two critical realities: firstly, we take much more into context than just delays. Secondly, an order that has absolutely zero contextual disruption but is simply exceptionally large must still trigger human oversight. In order to give the manager still the chance to approve it, as in the end he is still responsible for his gas station. Simultaneously, Eszter highlighted the need for clarity around the escalation numbers. This feedback forced a complete iteration of the platform’s core messaging and logic.
Therefore, I rewrote the “Station Decision Rule” to explicitly state that missing context data or exceptionally high values will still trigger manual reviews. This collaboration taught me that in B2B product development, peer feedback isn’t just about tweaking design; it fundamentally stress-tests the business logic. Furthermore, it helped me to understand that feedbacks are not only helpful, but also crucial as sometimes you are stucked with your view and “forget” about certain important layers.
My Role and the Iterative Crucible
My role in the assignment was to develop the prototype and think about how the AI could be used in MOL’s operational structure. Within generating the prototype one of the most important design decision I made was moving away from simple “Approve/Reject” modals. Instead, I architected the “Operational Decision” split. By forcing the system to explain why it was pausing an order (e.g., “26 items are ready, 1 is delayed”), and giving the manager (like our persona Marek Dvořák) the agency to release the available goods while backordering the rest, my contribution was shifting the prototype from a passive dashboard into an active, collaborative thought partner for the user. So for Marek his partner in crime.
Rethinking AI and Business Strategy
Before this project, I viewed AI in business strategy as a tool for cost-cutting and labor replacement. Something, also most people might think. Maybe also you, dear reader? Designing the ARDIP, our prototype, completely inverted that perspective. I now understand that the most effective AI strategies are rooted in Human-in-the-Loop architecture. Somehow this seem logically, however still as AI processes you might not think about exactly this. The legacy MOL Group system relied on a static rule (block everything over 12,500 Kč). Nevertheless, my initial instinct was to replace that static rule with a black-box AI algorithm as stated in the beginning.
But the true strategic breakthrough was designing a system that uses AI to gather hyper-local context (such as traffic data or supplier telemetry) and curates that data so a human can make a high-leverage decision within seconds. Marek would thank us for making his life easier as now he needs only couple of seconds to approve the order having all the informations he need, whilst in the earlier days it may take him maybe even hours.
Ultimately, this project taught me that the goal of AI in enterprise operations is not to remove the Station Manager from the equation. The goal is to remove the noise, so the manager can actually manage. Trust is the ultimate bottleneck to AI adoption, and trust is built through operational transparency—not algorithmic perfection. The core learning was that in supply chain operations, efficiency without transparency generates friction. True innovation happens when AI acts as a contextual advisor rather than an invisible dictator. Thanks for sticking till the end – it was really fun to generate a prototype.
Have a great day!
David Stjepanovic
