Jim Farley says Ford's next-generation electric vehicles will diagnose their own assembly errors, and the coverage treated it as a software story – clever car, clever sensors, another win for the telematics stack. I have stood at too many final audits, hold cage filling, truck at the dock, to read it that way. What I heard was an admission priced in advance: defects will leave the plant and be settled at dealer labour rates. The press heard a feature. I hear an escape metric with a marketing budget.

detection is not prevention

Occurrence and detection are separate axes in a PFMEA, and this industry has spent three decades quietly conflating them. An in-vehicle diagnostic capability improves detection – or at least the story people tell about detection. It does nothing for occurrence. The defect still forms at the same station, at the same rate, under the same tired eyes. But watch an FMEA review once such a system exists: the score falls, the risk colour turns green, the pressure to fix the forming process evaporates. The arithmetic does the work the process change was supposed to do.

There is a pattern here and it is not flattering. When catching defects gets cheap, making them stops hurting. End-of-line vision systems, leak-test booths, torque dataloggers – each arrived as a quality investment and functioned as a licence. No station team has ever got sloppier because a sensor downstream got better, officially. In practice the sensor buys the tolerance. And the escape ladder is unforgiving: an error costs €5 at its station, €80 at end of line, €600 at the dealer once the diagnostic hour, the trim teardown and the parts logistics land on it. Past that the currency stops being euros and becomes warranty claims and trust. Detection slides you up the ladder. Prevention removes the ladder.

A self-diagnosed defect is a confession, not a cure.

the escape ledger

A self-diagnosed error is still an escape. It crossed your last controllable point inside a vehicle you shipped, and the customer is now your inspection station – one that bills you.

At SNOP I built the quality function of a greenfield plant from the ground up, more than 900 people and no institutional history to hide behind. We cut defect costs 70%, and I can tell you where that number came from. Not smarter end-of-line detection. QRQC reacting at the point of defect formation, inside the shift, in the cheapest fix window anywhere in the value stream. The rule was blunt: the defect gets answered where it was born, not where it was found.

At Witte Automotive the A3 countermeasures and the Q-Wall taught the same lesson from the opposite direction. Once failure costs went onto a wall in euros, updated daily, nobody was comforted by detection anymore. Nobody asked how we would catch it. Everyone asked what had changed upstream – the only question that moves occurrence, and the one customers actually audit. In one quarter we carried zero critical customer escalations. Not one customer asked about our detection capability. They asked what we had changed so the defect could not form. Customers forgive we don't make it for years. We catch it invites an audit.

the data loop nobody staffs

Now the genuinely new part. A fleet that diagnoses itself is a distributed sensor network with millions of nodes, and the field data it produces will drown any human triage desk. I design and deploy multi-agent AI systems – one orchestration platform I built runs more than sixty models in parallel and synthesises their consensus – so let me say plainly which half of this problem is solved: triage. Clustering diagnostics, deduplicating repeats, ranking by severity and lot correlation, routing to the right owner. Software reads a thousand field reports in minutes. That is an engineering problem with an engineering answer.

Reacting is the other half, and software cannot do it. An 8D is five people, a containment decision, a supplier call and four weeks of discipline. Software does not sign a D5. The trade press is full of AI supplier-quality platforms and chatter about the coming ISO 9001 revision, and none of it answers the question that decides whether fleet diagnostics become gold or noise: who staffs the reaction? A million field findings arriving against a reaction capacity measured in dozens of open 8Ds is not a quality system. It is a queue with a dashboard.

Key takeaways

  • Score self-diagnosed field findings as escapes, priced at dealer labour and logistics rates – never as services. The vehicle left your plant; the invoice is yours.
  • In the PFMEA, let in-vehicle detection improve the detection ranking only. If anyone proposes to lower occurrence without a documented process change, refuse it in writing.
  • Wire the field-diagnostic feed into the QRQC clock – same time-box, same rule: the reaction lands at the station that formed the defect, not the dashboard that found it.
  • Automate the triage layer with agents now, and size the human reaction layer – 8D leaders, containment engineers, supplier quality – against the diagnostic volume before it arrives.

So count Farley's self-diagnosing cars for what they are: escapes with telemetry. Log them against the plant, run them on the QRQC clock, price them at the dealer rate, and leave the occurrence ranking exactly where it sits until someone changes a fixture, a torque strategy or a weld schedule. A car that makes its own diagnosis is not your quality system working. It is your failure arriving with better paperwork – and if that paperwork gets treated as a service, the stations upstream will happily keep writing it.