Real-time SPC with AI — statistical process control that reacts while you can still correct, not once you are already out.

The problem

SPC used to work with sampling and end-of-shift analysis. Today it is the black belt's bottleneck.

Classical SPC has earned its place in history — it disciplined industrial quality for half a century. But the way it is usually run on the plant floor has problems that operations feel every single day:

1. Sampling every N parts — between two samples plenty of drifting parts can go out without the chart ever noticing.
2. Retrospective review — the quality engineer looks at the chart at the end of the shift or the next morning and reconstructs what happened. The warning arrives after the fact.
3. Manual capture — the operator writes on a sheet or in a spreadsheet; by the time it reaches the statistical package there are timestamps squared up by hand and missing values that get imputed. The black belt spends half the project in SQL and pivot tables.
4. Western Electric rules living inside the engineer's head — the 8 classic rules are in the manual, but the one who really applies them is the person looking at the chart. If nobody looks, they are not applied.

The classical system (sampling + statistical analysis at end of shift) works — and it falls short the moment the cost of a contaminated batch far exceeds the cost of knowing earlier that it was drifting. Real-time SPC does not reinvent the methodology: it moves it to the moment when correction is still possible.

How it fits the IRIS system

iLEAN does not replace SPC — it turns it into a warm handover for the quality engineer.

Real-time SPC is what the Deming cycle (plan-do-check-act) always asked for: the loop closed at the speed of the line. iLEAN does not invent it — it closes it at a speed and with a consistency no single person could sustain. Connect is the filler that seals the cracks between the MES, the SCADA, the lab spreadsheet and the in-line Edge data.

The quality engineer no longer starts the shift reconstructing what happened. They start it deciding what to do with what the system already has ready. A warm handover for SPC.

The iLEAN pieces applied to real-time SPC:

• Edge — computer vision and/or in-line sensors, measurement on 100% of the parts, raw data in milliseconds. It works with no network.
• Connect — captures the complementary variables (SCADA, MES, IoT, the machine's old panel, the lab spreadsheet) and unifies them. Without that capture, real-time SPC is Edge-only statistics — useful but partial.
• Agents — the brain. They maintain Xbar-R, p, np and U charts as appropriate; apply the 8 Western Electric rules automatically; compute continuous CpK/PpK; filter out system noise; and alert the right manager through their own channel. They also detect sensor drift and flag when the MSA needs repeating.
• Three safety rings — any proposal that stops the line or adjusts the process goes through human validation. The chart flags it; the person decides and signs off.

See the full IRIS architecture →

Before and after

Sampled SPC with end-of-shift analysis vs. real-time SPC with iLEAN

Aspect	Sampled SPC, offline analysis	Real-time SPC with iLEAN
Coverage	1 part in every N, periodic	100% of the parts, continuous
Drift warning	At end of shift, when the chart is reviewed	The moment a Western Electric rule fires
CpK/PpK calculation	At batch close	Continuous, visible at any moment
Data capture	Manual or semi-manual, timestamps by hand	Automatic (Edge + Connect), consistent timestamps
MSA	Scheduled, no drift detection	Scheduled + drift detected by the agent, alert to repeat
Quality engineer's time	50% SQL/pivot tables, 50% judgement	10% judgement on clean data, 0% SQL

Impact estimate

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

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

• A line running classical SPC today (sampling + statistical analysis at end of shift), Cpk target around 1.33, scrap in the 1-3% range.
• Edge + Agents pilot on that line (100% measurement, live charts, automatic WE rules, MES/SCADA integration). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on how much weight quality carries in your cost and on the line's volume.
• Reduction in the quality engineer's data-cleaning time of the order of ≥30% (conservative estimate). Additional scrap reduction from earlier reaction.

And the black belt's reasonable doubt

«What if the AI fires alarms that aren't real?» — hallucination is a problem of free generation, not of anchored tasks. Applying a Western Electric rule to an Xbar-R chart is the most anchored task that exists in statistics: there is a formula, there is a data point, there is a binary decision. On this kind of task the best models sit below 1.5% error ^[1]. And even so, the alert reaches the manager as a suggestion, never as an action on the line. The three safety rings exist precisely for this.

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

You may also be interested in: Six Sigma with industrial AI · DMAIC with AI-assisted statistics · Poka-yoke with computer vision