Physical AI runs the line. Root cause still needs a witness.

Daedong is turning its Daegu plant into a "physical AI" hub by 2029, with robots and autonomous systems meant to lift tractor production capacity by 67%. Half the manufacturers I talk to have a version of that plan. The other half are drafting one. My discomfort with all of them is narrower than the usual jobs argument. Across two decades in automotive and aerospace, my fastest root causes never started with a control chart. They started with an operator saying the press sounded different on Tuesday. That sentence has cracked more defects for me than any dashboard I have commissioned. What happens to it when there is nobody left on the floor to ask?

What QRQC actually runs on

QRQC – quick response quality control – is sold as a method. It runs on testimony. You gather at the machine within the hour, and the engine of the exercise is the person who saw, heard or felt the process misbehave. The same holds for 8D and for five whys. Every "why" in the chain terminates at an observer. The method presumes someone standing at the end of every answer.

I ran QRQC and A3 at Witte Automotive long enough to take a substantial block of failure cost out of the business, and the case I remember best was not exotic. Intermittent burrs on a stamped latch component. Four days of 100% sorting at roughly €3,800 a day. The SPC charts sat flat, because burr height was not a controlled characteristic at that station – the data circled the defect without touching it. Then a woman who had run that press for fourteen years leaned in during the QRQC session and said it had sounded different since Tuesday. Like it was clearing its throat. The die had been changed Tuesday night, and the setter's shim stack sat 0.2 mm off the previous build. We re-shimmed, the burr disappeared, containment closed the same shift. Four days of data. One sentence from a witness.

The gemba walk is the same machinery in slow motion. You walk the line to harvest tacit signals – tone, smell, vibration, the way an operator hesitates half a second over a part. Dashboards average all of that away. Root cause rarely lives in the average.

Automation removes the witness before it removes the defect

Physical AI will push your mean-time-to-detection beyond anything a human shift achieved, and quietly wreck your mean-time-to-root-cause. Cameras and in-line metrology catch defects no operator ever would. But when the cause is a servo glitch, a gripper that micro-slipped for 200 ms at 03:40, or an agent that rerouted parts around a blocked buffer, the failure mode arrives as a log entry rather than an anecdote – and nobody sets retention windows for the moments that matter. High-frequency data gets sampled down or purged at 30 or 90 days, precisely because the anomaly that explains the defect lives at millisecond scale.

The modern 8D has a new failure mode: D4 terminating at a log file nobody kept. Storage is cheap. The investigation you can no longer run is not. Companies keep seven years of email for the lawyers and thirty days of process data for the engineers, which tells you exactly who management thinks it works for.

A log records what someone thought to record. A witness records what nobody thought to ask about.

Building a greenfield QA/QC department for 900-plus people taught me which control actually holds: escalation speed. We went a full quarter with zero critical customer escalations, and that was earned in hours of response, not paperwork. Empty the floor of witnesses and your response time stretches at the exact moment everything else accelerates.

Engineer the witness back in

The answer is not to slow the automation down. It is to treat the witness as an engineered capability, with requirements, owners and validation – the same status you give a critical dimension. In practice, four things:

• Stroke-level signatures. Force–displacement curves for every press stroke, torque profiles for every fastening, stored as traces. Not compressed into end-of-line averages.
• Timestamped traceability. Every part or lot mapped to the parameters of its birth: die, machine, coil batch, agent version, setpoint deltas.
• Anomaly-event retention. A ring buffer that freezes the minutes before and after any deviation to immutable storage, held for part life plus warranty.
• Acoustic and vibration baselines. The machine-readable version of "sounded different on Tuesday", so the ear that cracked my burr case survives the retirement of its owner.

Then write the escalation triggers into the control plan and PFMEA as validation requirements under IATF 16949 or AS9100, not as an IT afterthought. If evidence availability is auditable, it gets funded. If it lives on a wiki page, it gets deleted.

Most of my AI work is deploying and integrating autonomous multi-agent systems – I wire the agents together, I do not train models – and the first thing I specify is that every agent logs its own inputs, reasoning and decisions in an interrogable trail. An agent that cannot explain what it observed and chose has not replaced your witness. It has deleted her on purpose.

Daedong wants 67% more tractors out of Daegu by 2029. I hope they get them. But the winners of the physical AI decade will not be the plants with the most robots. They will be the ones whose quality system can still answer "what changed?" on the first morning of the empty floor. Done properly, automation compresses your response time. Done carelessly, it stretches it while the dashboards glow green – and the first you hear of that is an 8D with no D4.