X̄-R control charts with AI — drift shows up before the subgroup closes.

The problem

SPC does not fail because the statistics are bad — it fails because the data arrives late.

The X̄-R chart dates from the 1920s and it is still correct. What has expired is not the theory — it is the operation: the operator writes the measurement on a sheet, the technician types it into a spreadsheet at the end of the shift, the quality engineer looks at the chart on Friday. By then the drift has been producing scrap for three shifts, and the Western Electric rule fired 40 parts ago.

In multi-product plants it gets worse: subgroups mixed across SKUs, false alarms the team learns to ignore, and SPC ends up as decoration. The quality manager knows it — and distrusts it. Not the tool, but the chain that carries the measurement from the workstation to the chart.

The classic setup (sheet plus spreadsheet plus weekly review) works 99% of the time. That 1% is precisely when expensive scrap is about to be produced or when the OEM customer files a claim for an out-of-spec dimension. The whole point of SPC was to anticipate — and without in-line capture it arrives late almost by definition.

How it fits the IRIS system

iLEAN does not add another SPC package — it puts live data underneath the chart you already have.

The problem with SPC is not a lack of statistical software (there are good packages out there). It is that the data lives in islands: the digital gauge probe, the sheet at the workstation, the MES that does not speak to the probe, the quality manager's spreadsheet. iLEAN acts as the filler that seals those cracks — without asking you to change your SPC software or your MES.

Edge measures where there is vision. Connect carries the data where there is a digital gauge or a manual entry. The Agent computes the chart and evaluates the rules. The person decides whether to intervene.

The iLEAN pieces applied to X̄-R charts:

• Connect — captures every measurement at second zero with its context (dimension, product, operator, shift, machine). If a digital probe reads the dimension, it integrates directly. If an operator records it on a tablet, the Connect app captures it. If it only shows on the machine's old panel, a camera reads it. The measurement is not lost because it was "not integrated".
• Agents — build the correct subgroup (separating the active SKU), plot the X̄-R chart in streaming, evaluate the active rules (Western Electric, Nelson) and, when one fires, propose to the shift supervisor the most likely special cause for that pattern. The judgement to intervene still belongs to the manager.
• Edge — for dimensions measured by a camera (size, presence/absence, defect). Every part is a measurement, the subgroup builds itself, and the drift shows on the workstation tablet, not in Friday's meeting. Combined with the powder coating pattern (intervention before the oven), the SPC loop closes in milliseconds.

See the full IRIS architecture →

Before and after

Reactive SPC vs. live SPC with iLEAN

Aspect	Classic SPC (sheet + spreadsheet)	With iLEAN Connect + Agents
Data capture	Operator writes it down or exports at end of shift	Probe, Edge or tablet capture at second zero
Subgroup construction	Manual in a spreadsheet, mixes SKUs at changeover	Agent groups by active SKU automatically
Western Electric / Nelson rules	Watched by whoever looks at the chart on Friday	Evaluated on every new data point, immediate alert
False alarms	Routine, caused by mixed SKUs	Filtered by product context
Special-cause diagnosis	"We'll look at it next week"	Agent proposal to the shift supervisor on the channel they use
Auditor evidence file	Rebuilt from the spreadsheet	Chart + rules + corrective action with full lineage

Impact estimate

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

Block flagged as an estimate to be validated with concrete data from your plant.

• Multi-SKU plant with an SPC control plan on 4-8 critical dimensions, several SKU changeovers per shift, documented scrap and rework.
• Connect + Agent pilot on one characteristic dimension (the one that generates the most scrap). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on volume and unit cost.
• Defensible floor: reduction of scrap attributable to undetected drift of ≥30%. A single avoided OEM claim pays for the pilot.

And the CAIO's reasonable objection

"What if the Agent misclassifies a Western Electric rule?" — hallucination is a problem of free generation, not of anchored tasks. Evaluating a rule against measured data is an anchored task: the best models brought the error below 1.5% ^[1]. And even so, the decision to intervene stays human: the Agent proposes, the shift supervisor decides. The three safety rings exist precisely for that.

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

Related topics: Lean Manufacturing methods