The challenge
Our client's device controller worked well, but one issue was holding back growth: the device could not adapt to a new installation site on its own. A specialist had to configure each installation manually before commissioning, making the product difficult to scale through partners.
Every installation had to be configured manually
The controller monitors the device's power consumption and shuts it down if it detects a fault condition that could cause damage. However, the device behaves differently at each installation site. Load, device size, operating profile and disturbances in the electrical network all vary. A specialist therefore had to review the device's behaviour and manually determine when the controller should interpret the situation as a fault. The work took around ten minutes per installation, but required expertise that a typical field technician did not have.
Manual configuration was limiting sales growth
Partner distribution could potentially reach around one hundred new sites. But if every device still required specialist configuration, growth would remain constrained by expert availability. The bottleneck was not the product itself or market demand. It was the fact that every new installation required specialist support.
The most obvious AI solution could not be built
The initial idea was to train a model to determine the correct settings automatically from data collected at a new installation site. This would have required data from many different installations together with their known good settings. In practice, high-quality data was available from only one site. As a result, this approach could not be implemented reliably.
What we built
We ran a focused AI Design Sprint across five workshops. The goal was not to build a production-ready product, but to answer two questions quickly: can AI solve the problem, and what kind of solution should be productised? Between workshops, we analysed data, tested different models and refined the solution based on what we learned.
Results
The AI Design Sprint showed that specialist-dependent commissioning could be replaced with an approach suitable for a partner channel. At the same time, the client gained a clear path towards a 2026 product launch.
The machine learning model detected a dangerous fault condition with 92-94% accuracy using real operational data, and identified it within the first minute. This demonstrated that the problem could be solved technically and provided a foundation for further product development.
When specialist support is no longer required for every installation, partners can commission the device themselves. This brings around one hundred new sites within reach that a specialist-led operating model could not have served.
When a fault is detected early, a device replacement costing around EUR 10,000 can in some cases be avoided with a bearing and seal service costing around EUR 2,000. Earlier shutdown also significantly reduces the electricity wasted while the device is operating in a fault condition.
The AI Design Sprint produced a clear solution architecture and three concrete follow-up experiments. These provide a path to preparing the product for partner distribution in 2026 without having to solve everything at once. The solution can also be made more automated over time as more operational data becomes available.
