AI on the shop floor: where it earns its keep

Shop-floor AI has a credibility problem in smaller factories, and it is deserved. A decade of demonstrations involving robots nobody could afford has left a lot of owners assuming none of it applies at their scale.

Three applications now genuinely pay back in firms turning over single-digit millions. Each has a real prerequisite, and being honest about those is what separates a project that works from another dust-sheeted purchase.

Vision inspection

A camera and a model that decides whether a part is good. The economics have changed sharply: what needed bespoke engineering five years ago is now achievable with off-the-shelf hardware and a few hundred labelled images.

Where it pays: high-volume visual checks currently done by a person who is also doing three other things. Surface defects, missing features, incorrect assembly, label and orientation checks.

The prerequisite: you need examples of failures. If your scrap rate is so low that you cannot produce a hundred images of the defect, the system has nothing to learn from — and you may not have a problem worth solving.

The honest limit: it will catch the defects you trained it on. A genuinely novel failure mode passes straight through, which is why it augments rather than replaces final inspection on safety-critical work.

Predictive maintenance

Sensors on machines, watching for the signature that precedes a failure.

Where it pays: when a specific machine's unplanned downtime is genuinely expensive — a bottleneck, a long lead time on parts, or a failure that scraps work in progress.

The prerequisite: enough history to know what failure looks like. On a machine that has failed twice in five years, there is not enough signal to predict anything, and simple condition monitoring with sensible thresholds gets you most of the value for a fraction of the cost.

The honest limit: this is where the most money is wasted in SME manufacturing. Firms instrument twenty machines because a vendor sold a platform, when only two had downtime worth preventing. Start with the machine that hurts.

Scheduling and sequencing

The least glamorous and often the most valuable, because it needs no new hardware.

Where it pays: job-shop environments with many small orders, competing priorities and setup times that vary by sequence. If your production plan lives in a spreadsheet maintained by one person who is the only one who understands it, this is you.

The prerequisite: your routing and setup time data needs to bear some relationship to reality. This is where most firms discover their standards were set in 2011 and nobody updates them.

The honest limit: a schedule nobody follows is worthless. If the shop floor overrides the plan daily because it does not reflect how the place actually runs, fix that first — and it is usually a data problem, not a people problem.

What they have in common

All three depend on data you either already have or can collect cheaply, and all three are fundable. Made Smarter's match funding explicitly covers AI and machine learning, IoT and sensors — and the £20,000 cap, at 50% match, is a meaningful proportion of a well-scoped vision cell or a sensor package on the machine that matters.

Pick the one where you can already name the cost of the problem. If you cannot put a number on it, that is the work to do first.