How it started
When Priya first walked us through the plant, she showed us a whiteboard in the corner of the shift office. Three columns: running, wobbly, down. Every Monday morning, her team moved little magnets around to show which machines were where.
It was the most honest dashboard we'd ever seen. It was also where our work started.
The plant — a mid-sized precision-components manufacturer in the Midlands — was losing about 15% of its production time to unplanned stops. The team had the instincts to spot trouble a day or two early, but the institutional knowledge lived in four heads, and those four people couldn't be everywhere at once.
What we looked at first
Before we talked about AI, we spent a fortnight just watching.
We sat with the maintenance crew. We read a year of logs. We mapped which stops were genuinely surprising and which were, as one operator put it, "the same damn bearing on line three, again."
The answer, it turned out, was that about 70% of the unplanned downtime came from failure modes that were predictable — if you looked at the right sensor, at the right cadence, with a model that had seen enough of the pattern before.
What we built
Nothing exotic. A small, well-behaved system:
- Sensor layer: the plant already had 80% of the sensors we needed. We added a handful more around the worst-offending assets.
- A forecasting model trained per-machine, predicting failure probability over a 24-72 hour window.
- A maintenance dashboard that spoke plain English — "Pump 4 shows the same signature it had before last October's stop. Suggest inspection within 48h."
- An honest alert threshold, tuned alongside the team so they trusted it.
That last point mattered more than the model. Alerts that cry wolf get ignored. We ran the system in silent mode for six weeks before a single notification went to the maintenance lead, and we only turned it "on" once Priya said she believed it.
What changed
Twelve weeks in:
- 40% reduction in unplanned downtime, steady across the following two quarters.
- 25% decrease in maintenance spend, mostly from swapping emergency call-outs for planned work.
- ~£2.5M annualised savings on a project that cost a fraction of that to build and run.
"It's not magic. It's just that we finally hear what the machines have been trying to tell us for years."
— Priya, Operations Director
But our favourite number isn't on the dashboard. It's that the maintenance team took two full weekends off in a row for the first time in eighteen months.
What we'd do differently
Honestly? We'd start even slower. The temptation on any AI project is to build the model first and the trust second. In hindsight, every week we spent sitting in the break room listening paid itself back three times over in how well the final system was received.
If we'd skipped it, we'd have shipped a technically correct system that nobody used.
Where it's going
The plant has added three more production lines to the system, and we're helping them think about energy optimisation next — another place where small, consistent signals add up to big, quiet savings.
If you're looking at a similar problem — unplanned stops, institutional knowledge locked in a few heads, a team that would love a quieter Monday morning — let's have a chat. No deck, no sales pitch. Just a conversation.