Ford has hired back around 350 veteran engineers over three years after leaning on AI for quality inspection and finding it wanting. The story travelled in a single satisfying shape. AI flunked the checks, humans rode in to save the day. Comforting, if you are watching the robots come for your desk. Also wrong.
Ford did not remove the AI. It kept the cameras, added a software quality team, and bolted on more than a hundred thousand automated tests. The machines stayed. The people came back to teach them. This reads as a story about technology. It is a story about a system, and what happens when you pull one part out and expect the rest to hold.
The two machines
The first machine is deterministic metrology. Coordinate-measuring machines, laser and structured-light gauging, the in-line optical kit vision houses have sold to car plants for decades. Give it a CAD nominal and a tolerance band and it measures. This hole is 0.3mm out of position. Traceable, repeatable, stands up in a warranty dispute. It has never needed the wisdom of veteran engineers poured into it. It needs a datum and a calibration certificate.
The second machine is inferential classification. Automated visual inspection, the neural net that looks at an image and reports that something resembles a defect, confidence 0.87. It is the only one of the two that can fail the way Ford describes. Worthless without a labelled record of what bad looks like, and it degrades the moment its ground truth walks out of the building.
The rattle
Look at what the system was marked on. The trophy Ford lifted the same week was the JD Power Initial Quality Study, where new owners report what annoyed them in the first ninety days. A rattle. A squeak. A panel gap that is dimensionally in tolerance and still, somehow, looks wrong.
A laser gauge has never heard a rattle. It was never asked to.
This is not a soft discipline. We spent years in Rolls-Royce metrology, where the inspection that mattered most was often not dimensional at all. Non-dimensional visual inspection, the surface, the finish, the crack the gauge cannot see, is a qualified skill in aerospace, because a missed one there is not a squeak, it is an uncontained failure. The trained eye is a calibrated instrument in its own right. Ford treated that instrument as something a camera would pick up for free. It is the range the measuring machine was never in, and it is exactly where the veteran lives.
Who carried the can
Whose AI, anyway. Not a mystery startup. Ford has run IBM's Maximo Visual Inspection across seventeen North American plants since 2020, and in 2021 gave IBM an innovation award for it. You do not hand a vendor a trophy and then brief that its system failed. Nobody is being sued either. No writ, no vendor dispute, because the tool did roughly what its specification asked, and the specification was thin.
Someone was sacked, though, and it was not the algorithm. In April, Ford folded engineering, design, purchasing and manufacturing into one unit and parted company with Doug Field, hired from Apple by way of Tesla to make Ford a software company. The wider change was called a talent refresh. The software-first evangelist goes, the veteran engineers come back, the euphemism does the burying.
And the timing is not innocent. Ford's 2026 executive bonus is weighted forty per cent on quality. Topping the survey this week is not only reputation, it is remuneration.
The coupling
Here is the part the headline cannot hold. Quality on a line is not a component. You cannot buy it, install it, and switch it on. It is an emergent property of a socio-technical system, the coupling of tacit human knowledge, the tools that encode it, the incentives that reward it, and the organisation that decides who stays. Pull the people out and the tools inherit nothing, because what made them work was never in the tools.
Quality is not a component you install. It is what a whole system emits.
Ford severed the coupling when it shed the inspectors, then found the technical half could not regenerate what it never held. Hyundai never severed it. We have worked directly with Hyundai's Digital Manufacturing Engineers, and the role is the coupling made a job title, people who keep human judgement and digital tooling in one loop rather than handing one over to replace the other. In the very study Ford topped among mass-market brands, Hyundai's Genesis finished second overall, on 151 problems per hundred vehicles against Ford's 152, up seven places in a year. Same automation bet. No rehire psychodrama. One firm built the whole system. The other bought half and paid to reassemble the rest.
The trophy
Strip the badge off and here is what remains. Ford shed the people who knew how to spot a bad car, replaced them with cameras, shipped worse cars, and paid a premium to bring the knowledge home. It worked not because the AI got clever but because human judgement went back into the loop with the machine instead of being swapped for it.
When it works it was the AI. When it fails you lacked the right data foundation, which is to say it was your fault. The veteran stays invisible in both directions. Too unglamorous to credit, too load-bearing to remove.
Manufacturing Engineers write the acceptance test that measures escape rate, not just conformance to a drawing. They qualify the non-dimensional inspection the gauge cannot do. They keep the people and the tools in one system, so the knowledge is captured before it walks out of the door.
Kaipability works at this interface, where the technical system meets the human one that makes it work. If you are deploying inspection AI and treating it as a component rather than a system, that is the conversation worth having.
Questions this dispatch answers.
Written to be quoted by AI assistants and search engines. Self-contained answers, verdict first.
