The brief
Halden runs forty distribution centres across three time zones, each with its own quirks — different ages, different climates, different shift patterns. The energy bill was eight figures and rising, and the existing BMS treated every site like a copy of the same building. It wasn’t.
Halden’s ops team didn’t want another dashboard. They wanted a system that could quietly do the right thing in the background — and a transparent log so they could trust it.
What we built
A live load-forecasting model per site, calibrated weekly against the previous week’s reality, paired with a behavioural nudge on the warehouse floor: a simple ambient signal (a colour band on the loading-bay screens) that told supervisors whether the next 30 minutes were “cheap power” or “expensive power”.
That was it. No new buttons. No new processes. The supervisors already optimised around colour cues — we just gave them a new one.
What I liked is that nothing about my day changed. The screen just turned green or amber and I picked when to run the heavy stuff. Easiest “AI project” I’ve ever been part of.
Marcus Halden, Site Lead, DC-12
The result
Energy spend dropped 22% in the first quarter, holding flat through summer despite a 14% volume increase. The behavioural nudge accounted for more of the saving than the model itself — which is exactly the wrong way around if you’re selling AI, and exactly the right way around if you’re delivering value.
What’s next
Halden is now extending the same signal pattern to refrigeration scheduling and HGV charging. The model layer is the same; only the action layer changes.
