The collaboration originated from a process challenge that had never been solved computationally: when exactly should operators switch steam pressure levels during crystallization? The decision — moving from step 4 to step 5 and back — determines energy efficiency across the entire batch. It was made entirely through decades of tacit, sidemand-learned expertise. No documentation, no transferable procedure.

NeuroSpace named the resulting system OWLS — Optimal Vacuum Vapour Level Setpoint — coined by two Harry Potter fans on the team who wanted a name that stuck. But before any model was built, the team did something equally important: they put on disposable overalls, stood on a milk crate on the factory floor, and introduced themselves to the operators. “We’re not here to replace you. We’re here to see if we can increase the quality of what you produce.” One operator — 30 years on the job — responded: “So what you’re trying to do is increase our collective piecework rate?” The answer was yes. His reply: “Welcome.”

That moment of trust, NeuroSpace argues, is what made the project succeed. You can build the best model in the world — but if operators don’t adopt it, nothing changes.

The numbers tell the story clearly. Before the system, operators made approximately 30 manual adjustments per campaign. In the first period after introduction, they made 80 out of a possible 124 — near-total adoption. Over time, as the predictions became familiar and routine, usage naturally declined. As NeuroSpace noted: “They got used to them” — a sign not of failure, but of the system becoming absorbed into normal judgment.

In 34 days of operation, the system achieved 39.8% steam efficiency improvement — reported conservatively to avoid greenwashing accusations, though the actual figure was 40%. The theoretical range spans 7% (conservative) to 43% (aggressive), depending entirely on how operators choose to act on the forecast. That range is not a flaw in the model. It is the HI design: the machine predicts, the human decides. Total direct savings in the initial run: 3,000 liters of diesel across the vacuum heating units, with 26 units in Nakskov and 26–28 in Nykøbing Falster.

NeuroSpace’s standing principle captures the dynamic precisely: “Every time we exclude domain experts, we fail.” The model that achieved these results was only possible because the team first embedded themselves in the process. The AI did not replace that expertise — it encoded and scaled it.

The Nordic Sugar case illustrates what HI looks like when it works at industrial scale: a human-AI system where the machine handles prediction and the operator retains full decision authority — and where the value unlocked depends entirely on how well tacit human knowledge has been built into the model. For organizations in energy-intensive industries, the case offers both a concrete benchmark and a replicable methodology: embed domain experts in the design process, build for the operator’s actual workflow, and measure success not by model accuracy but by whether human decisions improve. The project has now been running for three years — a testament to what happens when hybrid intelligence is designed to complement rather than replace the people who know the process best.