Standard work with AI vision — the standard stops being a laminated sheet on the wall and comes back to the station, alive.

The problem

The standard is written, laminated and hanging on the wall. And nobody has read it since day two.

Standard work is one of lean's best ideas: fix the best-known sequence so that quality does not depend on who happens to be at the station that day. The idea is still good. What fails is the medium — a laminated sheet on the wall — and it always fails in the same four places:

1. Nobody looks at it — the veteran has it in their hands and does not need to look; the new operator looks at it for two days and then copies whoever is next to them, which may or may not match the written standard.
2. The deviation is discovered late — at final inspection, in the customer claim or in the process audit. That is, when parts have already been made, cost is already sunk and an investigation is reconstructing after the fact something nobody recorded.
3. The standard ages on the wall — the team improves the method in a kaizen, the improvement stays in three people's heads and the laminate keeps describing how the work was done two years ago. Updating it means drafting, approving, printing, laminating and replacing it at every station: days. At that cost, it almost never happens.
4. The drift is silent — nobody decides one Tuesday to skip a step. Steps get lost through an accumulation of small, reasonable shortcuts, and each one on its own looks harmless until the defect appears that can only be explained by combining three of them.

The result is a paradox familiar to any certified plant: the standard is impeccably documented for the audit and, at the same time, nobody can state with data that it is being executed. Documenting the standard and verifying the standard are two different problems, and until now we had only solved the first one.

How it fits the IRIS system

iLEAN does not rewrite your standard — it puts it back at the station and checks that it is followed.

Standard work always asked for two things that in practice never held up: that the instruction live where the work happens, and that someone check continuously that the real sequence matches the standard one. iLEAN does not change the methodology: it gives it the support it was missing. The instruction lives on the station tablet and updates itself; the checking is done by Edge on every cycle, without tiring and without interrupting anyone. Connect is the putty that fills the cracks between the MES, the smart tooling and what the camera sees.

The process is verified, not the person. If a step is systematically skipped at a station, the problem is almost never the person executing it: it is the standard, the tooling or the instruction. That is exactly the finding we are after.

The iLEAN pieces applied to standard work:

• Edge — a camera over the station recognizes process events (part present, tool picked up, component inserted, tightening completed) and orders them on a timeline. All processing happens at the station itself, and what leaves the terminal are events, not images of people. It works without a network.
• Connect — brings the signals vision cannot see on its own: the work order in progress, part number and variant, screwdriver torque, weight, readings from smart tooling. With this the system knows which standard applies on each cycle, not just whether the cycle resembles one.
• Agents — compare the observed sequence against the standard sequence, classify the deviation (skipped step, altered order, time outside the window), filter out the operation's normal noise and alert in the moment through the area lead's channel. They also aggregate by station and by part number, so the team sees where the standard meets resistance.
• Living instruction — the standard rendered step by step on the station tablet, with the correct variant loaded automatically. When the team improves the method, it is published and is at every station that same shift: knowledge SMED.
• Three safety rings — no deviation stops a line by the machine's decision. The system proposes; the responsible person decides and signs.

See the full IRIS architecture →

Before and after

Laminated standard vs. standard work with AI vision

Aspect	Laminated standard on the wall	Standard work with AI vision
Where the instruction lives	Sheet on the wall, the same for every part number	Tablet at the station, with the current order's variant
Compliance verification	Sampled: a process audit every X weeks	Continuous: compared against the standard sequence on every cycle
When the deviation appears	At final inspection, the claim or the audit	In the moment, while one sentence still fixes it
Updating the standard	Draft, approve, print, laminate, replace: days	Edit and publish: the same shift, versioned and signed
Training a new operator	Constant shadowing by a veteran, variable duration	Living instruction + sequence confirmation; the veteran steps in where needed
Unit of analysis	Nonexistent, or noted by hand during the audit	Aggregated by station and part number — never by individual
Audit traceability	Reconstructed after the fact	Version and compliance history, already prepared

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We set it out so the committee has an order of magnitude; we refine it during the diagnostic.

• Manual assembly station or cell with several part numbers, a standard documented on paper and sequence defects that today are caught at final inspection or downstream.
• Edge pilot on 1-2 representative stations: the standard loaded as a sequence, living instruction on the tablet and comparison on every cycle. First value expected within a few weeks.
• Indicative payback between 5 and 10 months, resting on two levers: fewer sequence defects (less rework, less scrap, fewer claims) and faster training of new operators, which frees up veteran staff hours.
• Reduction in the time a new operator takes to reach the veteran's quality in the order of ≥30% (a conservative estimate), and standard update time down from days to minutes.

And the production manager's reasonable doubt

“What if the system flags deviations that are not deviations, and we end up with noise at the station?” — it is the right objection, and the answer lies in the type of task. Hallucination is a problem of free generation, not of anchored tasks. Comparing an observed sequence against a declared standard sequence is a textbook anchored task: there is a list of steps, an expected order, a time window and a deterministic comparison. In this class of tasks the best models are below 1.5% error ^[1]. And where a specific check does not reach sufficient reliability, it is not forced: it is flagged as non-verifiable and generates no alert. We prefer a standard covered at 85% with solid signals to one at 100% with false alerts the station learns to ignore in two weeks.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.

You may also be interested in: Operator skills matrix with AI · Poka-yoke with machine vision · Yamazumi and line balancing with AI