I've spent time in a lot of factories. Some noisy, some eerily quiet. What I've noticed across all of them is that the best plant managers run on a particular, hard-won kind of calm. They've seen things go wrong before, and they trust their team to handle what comes.
AI, at its best, adds to that calm. At its worst, it interrupts it with notifications.
Here's my attempt at a warm, practical tour of what AI is really doing in manufacturing today, from conversations with plant managers, maintenance engineers, and line operators — not consultants.
The unglamorous hits
Predictive maintenance. The case study of our work with a Midlands manufacturer is elsewhere on this site — the short version is that sensors plus a forecasting model cut unplanned downtime by 40%. That's a big number, but it's not even the most important outcome. The maintenance team got their weekends back. That's the outcome.
Quality inspection. Cameras above the line spotting defects that human inspectors miss, especially at the end of a long shift. The best implementations don't replace the human inspector — they give her a second opinion. She still makes the call. She just catches more.
Demand forecasting. Modern time-series models plus weather data, calendar data, and order history produce forecasts that genuinely beat the old methods. Not by a revolution — by 5-15%. In low-margin manufacturing, that's the difference between a good year and a great one.
Energy optimisation. Factories are voracious energy consumers, and small, consistent optimisations add up. An AI scheduler that slightly reshuffles when high-load machines run (to line up with cheaper or greener grid hours) can cut energy costs by 10%+ with no capital expenditure. This one's close to my heart.
Safety. Vision systems that spot when a worker enters a dangerous zone without PPE, or when a forklift is moving too fast in a pedestrian area. Done well, these systems warn the worker and log the event — without surveilling in a way that feels intrusive.
The less glamorous truths
Implementation is 80% of the work. The model is rarely the hard part. Integrating with legacy PLCs, the SCADA system, the MES, and the ageing ERP — that's the hard part. Budget accordingly.
Sensor hygiene matters more than model choice. If the sensor data is noisy, miscalibrated, or inconsistent, no model will save you. A couple of weeks of careful data work upfront is usually the best ROI on any AI-in-manufacturing project.
Operators must be on your side. The plant's wisdom lives in the heads of the people who work there. If they don't trust the AI tool — because it's been imposed on them, or because it cries wolf — they'll find ways to ignore it. A week of workshops, with the operators as equal participants, changes everything.
Legacy is real. Many factories are running control systems older than some of the engineers who maintain them. "Let's add an AI layer" in this context often means "let's add a modern read-only observability layer that doesn't touch the critical path." That's usually the right answer.
What great looks like
A plant manager once described to me what a successful AI deployment felt like, a year in. "It's quieter," she said. "The alarms that used to go off at 3am don't. The engineers aren't running around patching things that broke overnight. My Mondays aren't triage any more. They're planning."
That's the standard. Not transformation. Calm.
Where I'd start if I were you
For a manufacturer considering AI for the first time, here's what we usually recommend:
- Pick one asset, one failure mode, one shift. The narrowest possible scope. Ship it. Learn.
- Measure against a specific baseline. "Unplanned stops on line 3 due to bearing failures, month-on-month." Not "efficiency."
- Involve the maintenance team from day one. They'll tell you what's worth building. They'll also be the ones who make it stick.
- Plan for the handover. The system needs to be maintainable by someone who isn't the consultancy that built it. Write that into the contract.
What's coming
I'm watching a few quiet trends:
- Foundation models for industrial telemetry. General-purpose time-series models that can be fine-tuned on a specific plant's data with far less effort than bespoke models used to take.
- Digital twins that are actually useful. The first generation was mostly marketing. The current generation is starting to pay its keep for process simulation and operator training.
- Vision-language models for operator support. "Point your phone at this machine and ask a question" — and get a correct, grounded answer. We're close on this, and it's going to be big for maintenance workflows.
If you're thinking about any of this — or just want to trade notes on what's working on your floor — we'd love to chat. Factory conversations are some of our favourites.
