BMW deployed a humanoid. Nobody wrote the PFMEA for it.

BMW put the Figure 03 humanoid on the production line at Spartanburg. I've read the coverage. Every headline asks the same thing: does it work, can it grip, can it hold takt. Nobody is asking the question that actually matters when a semi-autonomous physical agent enters a controlled manufacturing process. Does the quality system around it work?

Same week BMW was celebrating, Ford was quietly rehiring engineers it had shed when it replaced them with AI quality control. Three hundred fifty of them. The system couldn't hold the standard. That's not an AI failure — it's a quality architecture failure. The wrapper around the technology, not the technology itself. And it's a preview of what happens when you deploy adaptive systems without the discipline to contain them.

The variation you didn't characterise

Traditional automation is deterministic. You validate a robot cell, prove the cycle time, confirm positional repeatability to the micron. Then it does the same thing two million times. Your PFMEA accounts for mechanical wear, sensor drift, end-of-arm tooling failure. You know the failure modes because they're fixed. They don't wake up one morning and decide to do the job differently.

A humanoid does exactly that. Figure 03 and systems like it are designed to adapt. That's the entire pitch — they handle parts that aren't perfectly placed, fixtures that sit slightly off, lighting that shifts through the afternoon. But adaptation is learning, and learning inside a controlled process is drift you didn't gate. AS9100 and IATF 16949 live or die in that gap.

When a human operator develops a workaround — and they do, constantly — a team leader catches it on a gemba walk or a layered process audit. The workaround gets documented. Standard work gets updated. The PFMEA gets reviewed. Loop closed. When a humanoid quietly optimises its grasp pattern because its objective function found a marginally better path, who catches that? The behaviour is embedded in parameter shifts invisible on the shop floor. Your control plan says the robot performs step seven a specific way. The robot has decided otherwise. Your control plan now describes a process that no longer exists.

The control plan you wrote describes the process you validated. The process running on your line is the one the robot learned instead.

Stop authority for something that learns

Every plant I've operated has had a clear stop-authority architecture. Operators pull the andon. Team leaders escalate. QRQC triggers within the hour for quality escapes. The 8D opens. A3 problem-solving runs its course. This chain works because it assumes two things: the operator or machine at the station is executing the validated process, and when they deviate, the deviation is visible to someone who can act.

A humanoid breaks both assumptions at once. Its learned behaviour can diverge from the validated job instruction with no external signal. The cell looks fine. Cycle time holds. The part passes in-line gauging because the defect the robot introduced is process-level, not part-level — not yet. You've built an escalation system for failure modes you can see. The humanoid's failure modes are behavioural, incremental, and buried inside a model you didn't train and can't inspect at station level.

I've built QRQC systems that close quality escapes inside 60 minutes. That system depends on a characterised failure mode and a reaction plan that already exists. You cannot run QRQC on a defect mechanism you haven't identified. And the fundamental value proposition of adaptive AI in manufacturing is that the system will develop behaviours you didn't identify and can't predict. That's not an edge case. That's the design intent.

What aerospace does differently

In my role as Head of Manufacturing Engineering Technical Authority for Airbus in North America, I gate new manufacturing technologies into a controlled aerospace environment. Every new process — every tool, every method, every piece of equipment — gets a PFMEA revision, a control plan update, and first-article validation before it touches a deliverable. We don't deploy first and characterise later. That sequence is not optional, and it's not bureaucracy.

That discipline is why we dropped EASA audit findings by 50% in one cycle, and why external audits run consistently clean. When you introduce a new manufacturing technology in an AS9100 environment, the framework forces you to answer questions automotive is currently skating past. What are the failure modes of an adaptive process. How you detect behavioural drift when the agent's decision logic is opaque to the operator. What happens when the validated process and the executed process diverge without an alarm — and who holds stop authority over a machine that has modified its own behaviour since the last audit.

If you can't answer those, you haven't introduced a new technology. You've introduced an uncontrolled process. Uncontrolled processes are the most expensive things in manufacturing. They just don't show up on the cost line until the customer finds them.

Humanoid robots will get more capable. Figure 03 is version one of something that will move faster, grip better, adapt further with each iteration. But the plants that succeed with these systems won't be the ones that deployed first. They'll be the ones whose PFMEA already had a line item for a process that changes itself. Whose control plans accounted for behavioural drift. Whose stop-authority architecture was built for a world where the machine on the line is learning faster than the quality system that governs it. Ford just learned that lesson with 350 rehired engineers. BMW is about to learn it on a production line with a humanoid nobody wrote the failure modes for.