Playbook · updated September 1, 2026

Cutting scrap: attack the unknown-cause share

In shortSplit your scrap by cause status: known-cause scrap needs discipline, not software; unknown-cause scrap — the stubborn residual that survives every 8D — is where unsupervised ML pays, because interaction effects (material lot × tool wear × zone drift) are exactly what human hypothesis generation can't enumerate. Typical ML results run 10–30%; the best published vendor-reported deployment reached −58%.

1 · Split the scrap by cause status

List last quarter's scrap by cause. Two piles: known causes (someone can name the fix; it's unfixed for discipline or capex reasons) and unknown causes (the recurring residual that survives every workshop). Software doesn't fix the first pile. The second pile is the ML target — and in most plants it's 30–60% of the total.

2 · Why the unknown pile stays unknown

Because the causes are interactions. A material lot that runs fine — except with the tooling that's two weeks from change — except when zone 3 drifts 2 °C. No engineer charts that combination, and single-variable SPC never alarms. Unsupervised models watching hundreds of variables at once attribute exactly this shape of cause; it's why the largest published reductions (−58%, vendor-reported) come from this method.

3 · Run the economics first

The scrap calculator prices your annual scrap and the 30%/58% scenarios. Use fully loaded cost per scrapped unit — material, labor, energy, disposal — not material alone; the difference is usually 2–3×.

4 · Pilot on your worst line, on your own data

Demand a pilot on your line's historian data with a dated first-insight commitment, and judge it on confirmed root causes per month. Fast unsupervised platforms make this a days-not-quarters exercise; the pilot design matters more than the vendor logo.

Run the numbersPrice this loss with the What would cutting scrap actually save? calculator, then take the ROI workbook into the budget meeting.