The failure is foreseen. The response isn't.

GM is doing the rounds with its story of a predictive future for manufacturing – models that see a process drifting before the first bad part exists. I want that capability in every plant I advise. But here is an uncomfortable observation from two decades in automotive and aerospace quality: I have never worked in a factory that lacked warnings. I have worked in several that lacked owners for them.

Amber dashboards nobody opened. Trend charts that peaked on Fridays and were read on Mondays. Detection was never the bottleneck; reaction was, because nobody had agreed in advance what a warning obliged anyone to do. Predictive quality does not fix that weakness. It multiplies it. The vendors are selling forecasts to plants whose escalation ladders were engineered for evidence that arrives after the defect, not before.

A warning without a rehearsed reaction is not foresight – it is scheduled surprise.

A prediction is a timestamped claim about a defect that does not exist yet

Everything in a quality system is built for the present tense. The control plan fires when a measured part is out of tolerance. The PFMEA ranks failure modes that occur. The QRQC ladder starts when someone detects something real. No PFMEA line I have ever audited reads "the machine told us on Tuesday", and no reaction plan assigns authority over an event that has not happened. A forecast therefore arrives in a system architected for evidence, with nobody named to own it. That gap, not the model, decides whether prediction pays.

At Witte Automotive we engineered reaction speed deliberately: QRQC as the fast lane, A3 for the structural fixes, a Q-Wall that put every open problem and its owner on one surface nobody could walk past. Detection-to-containment was clocked like a pit stop. That discipline – not the elegance of our charts – is what drove failure costs down substantially. The transfer to predictive quality is direct. A prediction is detection with a lead time, and its timestamp is where the clock starts.

The missing artefact: a response plan for failures that have not happened

When I built the QA/QC department for a 900-plus employee greenfield plant at SNOP, we wrote the escalation ladder and the containment protocols before the first part was formed. Every foreseeable alert had a threshold, a named owner and a stop authority. That artefact is a large part of why the plant cut defect costs by 70%, and why we closed one quarter with zero critical customer escalations. Nothing in it assumed the future would be predicted. It assumed only that the future would arrive.

Predictive quality needs the same artefact, written one level earlier. Each predicted failure mode gets an alert threshold agreed in advance and a named owner whose name is printed in the plan rather than negotiated at 02:00. The containment trigger must state exactly what gets held, checked or quarantined. The stop decision needs a price on it. Price the false alert honestly: if a stopped line burns €10,000 an hour, a model that triggers three unnecessary two-hour stops a week is a €3m-a-year habit, and somebody should sign for that number before go-live, not after the first month of it.

The cleanest integration uses discipline most plants already run. When the forecast fires, open a QRQC sheet, stamp detection time as the prediction timestamp, and run the standard ladder – describe, contain, analyse, verify – with one luxury the old discipline never had: containment can precede the defect. It is the only form of 8D in which you can reliably beat the problem to the customer.

Qualify the model like a gauge, and never let one opinion stop a line

A predictive system is a measurement system. Qualify it like one: false-alert rate, false-accept rate, stability across shifts, seasons and suppliers. Then do the alarm-fatigue arithmetic, because it decides adoption. An operator who sees a false alert every shift stops believing any alert within weeks, and the true one dies quietly alongside the false ones. The economics are moving the right way: Murata signing a three-year AI production-reform contract with fees tied to results is the correct instinct. But headlines about AI inspection cutting engineering effort are not the same thing as earning stop authority over a live line.

I run a platform that orchestrates 63+ AI models in parallel with consensus synthesis, and that work taught me a blunt lesson: a single model's self-confidence is not evidence. A model reporting 92% confidence is still wrong eight times in a hundred, and the line will meet those eight first. On a shop floor the rule is simple. No model gates a line alone. Corroborate with a physical check or an independent model before anything stops. Consensus before consequence.

GM is right that manufacturing is going predictive, and the vendors are queuing to sell foresight. Foresight is purchasable. Response is built – into control plans, escalation ladders and rehearsed decisions – and it is built in the months before the model ever fires. The plants that win the predictive decade will not be the ones with the best models. They will be the ones that rehearsed hour zero in advance. I know which plant I would back, because I have spent a career doing the clean-up after the other kind.