AI in manufacturing and operations: safe, lawful and under control

Cue 1 A production line with four AI uses marked: a camera, a vibration sensor, a schedule and a collaborative robot

EMMA 00:00 George, we've just had a camera system put on line three to inspect parts. Everyone's delighted. Should I be?

GEORGE 00:07 Probably, with your eyes open. Let me set out where AI turns up on a shop floor. Visual inspection, like yours. Predictive maintenance, watching vibration and temperature. Planning and scheduling. And robots and guided vehicles that work near people.

EMMA 00:23 Are they all the same risk?

GEORGE 00:25 No. The first three give advice or decide about product. The fourth moves in the same space as a person. That difference decides how much proof you need before you trust it.

Cue 2 Two bins, one of good parts wrongly scrapped and one of bad parts wrongly passed, with the second reaching a customer

EMMA 00:38 So what can go wrong with my camera?

GEORGE 00:41 It can be wrong in two directions. A false reject scraps a good part. That costs money and you notice quickly. A false accept passes a bad part. That costs far more, and you may not find out until a customer does.

EMMA 00:58 The supplier told me it's ninety nine per cent accurate.

GEORGE 01:02 Then ask the better question. How often does it pass a defect, on which defects, and how do we know? A system tuned to reduce scrap will pass more bad parts. Ask for the false accept rate on the defects your customer cares about most.

Cue 3 A graph of accuracy falling quietly after a change of material, with no alarm

EMMA 01:20 And how would I know?

GEORGE 01:22 Only one way. Keep testing it against parts whose true state you've established some other way. Known good, known bad, fed through regularly.

EMMA 01:31 It's been perfect for a month.

GEORGE 01:34 Which is where drift comes in. The system learned from data gathered at one time, under one set of conditions. When conditions change, what it sees no longer matches what it learned from, and its accuracy falls.

Cue 4 A list of everyday causes of drift around a camera and a part

EMMA 01:48 Wouldn't it tell me?

GEORGE 01:50 No. Nothing breaks. No alarm. That's what makes it dangerous.

EMMA 01:54 What sort of change?

GEORGE 01:56 Ordinary ones. A new material or supplier. A different surface finish. Someone changes a light fitting or the lens gets dirty. A tool wears. A new product variant. A sensor swapped for a different model.

Cue 5 A management of change form with four triggers: install, retrain, update, process change

EMMA 02:10 We changed steel supplier last week.

GEORGE 02:12 Then that's a change the inspection system should have been checked against. Which brings me to the law.

EMMA 02:19 Is there a law on AI in factories?

GEORGE 02:22 The Health and Safety Executive says existing law applies. The 1974 Act is goal setting, so it covers a risk whatever technology creates it. They expect a risk assessment for uses of AI that affect health and safety, and controls so far as is reasonably practicable. Including against cyber threats.

EMMA 02:42 So nothing special. Just do it properly.

GEORGE 02:45 That's their stated aim. That AI risk stops being novel and is managed like any other. In practice, management of change. Putting the system in is a change. Retraining it is a change. Updating its software is a change. And changing the process it watches is a change.

EMMA 03:04 And whose job is that? The supplier's?

GEORGE 03:07 Yours. The duty holder is the employer. Buying the system doesn't transfer the duty to assess and control the risk it creates.

Cue 1 Two screens: one advising an operator who decides, one acting on the line by itself

EMMA 03:16 George, last time you told me the camera can go quietly wrong. So how do I stay in charge of it?

GEORGE 03:24 Start with one question about any AI system on your site. Does it advise, or does it control?

EMMA 03:32 What's the difference in practice?

GEORGE 03:34 A system that advises shows a person something and the person acts. A system that controls acts itself. Rejects the part, stops the line, changes a setpoint.

Cue 2 A validation record beside a regular tray of known good and known bad challenge parts

EMMA 03:44 Ours rejects parts automatically.

GEORGE 03:46 Then it controls, for quality. That's a decision someone should have made deliberately. Anything new should start as advice. Moving it to control needs its own assessment. And where the action affects safety, it's part of the safety system and has to be designed and validated as one, by competent people.

EMMA 04:06 What usually goes wrong?

GEORGE 04:08 Not a bad decision to automate. It's an advisory tool that drifts into control because people stop checking it. If the operator always accepts what the screen says, it's controlling, whatever the procedure calls it.

Cue 3 An override log with a rising line marked drift and a flat zero line marked nobody is looking

EMMA 04:22 So how do I keep it honest?

GEORGE 04:25 Validate before use and watch in use. Before you rely on it, test it against cases whose true answer you know. Include the hard ones and the defects that matter most. Record the result, the conditions, and who signed it off.

EMMA 04:41 And afterwards?

GEORGE 04:42 Challenge samples. Known good and known bad, fed through at a set interval. And track overrides. How often people overrule it, and why.

Cue 4 An AI system on the plant network with a line out to a supplier and three questions beside it

EMMA 04:51 What's a good override rate?

GEORGE 04:53 It should never be zero. A rising rate is an early sign of drift. But a rate of zero means nobody's looking. And keep the manual method alive. If the team's forgotten how to inspect by hand, you've no fallback.

EMMA 05:09 Our IT manager is nervous about the supplier's remote access.

GEORGE 05:13 Rightly. It's software connected to plant, usually with a route back to a supplier for updates. That makes it part of your cyber security problem. The HSE names cyber threats explicitly, and manufacturers of things like machinery and vehicles are among the sectors the EU's NIS2 Directive covers.

Cue 5 A calendar marking 20 January 2027 and 2 August 2028, and four records in a folder

EMMA 05:32 What should I ask?

GEORGE 05:34 How is it updated and by whom? What can it reach on the network? And what happens to the line if it's unavailable, or gives wrong answers on purpose?

EMMA 05:46 We also build machines and sell some into Europe. Anything coming?

GEORGE 05:50 Two things. The EU Machinery Regulation applies from 20 January 2027 and replaces the old Directive. It addresses machinery whose safety functions rely on systems that learn. And the EU AI Act treats AI used as a safety component of a regulated product as high risk. After the July 2026 amendment, that applies from 2 August 2028.

EMMA 06:13 Is that UK law?

GEORGE 06:14 Neither is, and Great Britain has its own machinery regulations. If you export, take advice on your own product. What I've given you is direction, not a compliance route.

EMMA 06:26 And what do I keep on file?

GEORGE 06:29 Four records. What the system's allowed to do. How it was validated. How it's being checked. And every change made to it.