In factories, energy and logistics, AI is delivering, but it looks almost nothing like the chatbot economy that dominates the headlines. There are no witty assistants here. Most of the value comes from classical machine learning on sensor data: regression, anomaly detection, computer vision and time-series forecasting models that have been maturing for a decade. That's not a sign manufacturing is behind. It's a sign manufacturing already figured out where the money is.
Where the gains are
The wins cluster around three repeatable patterns, each tied to a physical cost that's easy to put a number on:
- Predictive maintenance. Anomaly detection on equipment sensors flags a bearing or motor that's drifting out of spec, so it gets serviced during planned downtime instead of failing mid-shift. Unplanned downtime is one of the most expensive things that can happen on a line, which is exactly why a model that buys you warning is worth so much.
- Quality inspection. Computer vision catches surface defects, misalignments and contamination faster and more consistently than a tired human eye on hour seven of a shift. The camera never blinks and never has a bad day.
- Optimisation of scheduling, routing, batch sizing and energy use, where a small percentage gain becomes a large absolute saving once you multiply it across thousands of units or megawatt-hours, the same territory our piece on whether AI can fix the energy problem it creates covers from the grid side.
Why this matters
The reason classical models dominate here is that the problems are narrow, repetitive and measurable, and the data already exists. A line that's been running for years has produced millions of sensor readings, each implicitly labelled by what happened next. That's an ideal setting for the boring, reliable models that generative AI overshadows in the press but rarely beats on the factory floor.
It also means the ROI conversation is unusually honest. You're not chasing a vague "productivity uplift"; you're comparing the cost of the model against the cost of a single avoided outage. (See measuring AI ROI for why that grounding matters everywhere, not just in factories.)
Why generative AI is slower here
The bottleneck is physical and infrastructural. Software moves only as fast as the hardware it observes, and the cost of a wrong action on a production line (a halted robot, a scrapped batch, an injured worker) is high. So autonomy stays tightly bounded by design.
In the field, AI earns trust the slow way: by being measurably right on a narrow task, then expanding. There's no demo shortcut, and the people signing off have seen too many demos.
What this means for an engineering team
If you're building for an industrial setting, resist the pull toward the flashiest model:
- Start where there's already a labelled history and a clear cost of being wrong. Predictive maintenance and inspection are the classic beachheads.
- Treat the model as a sensor, not a decision-maker: it raises a flag, a human or a tightly-scoped rule acts on it, often a small language model tuned to the one narrow signal it needs to catch.
- Budget for integration with PLCs, SCADA and MES systems, which is usually harder than the model itself.
The newest layer is generative on top: copilots that let an engineer query sensor history in plain language, or summarise an incident report into a maintenance ticket. That's genuinely useful, and it's where the next wave of value sits. But the engine underneath remains classical and quietly effective, and it's likely to stay that way for years. The smart move is to let generative AI make the proven systems easier to use, not to replace them with something that demos better and trusts less. If you're building for industrial and energy operations, that's a conversation worth having with our team.