Ford Called Back 350 Engineers After AI QA Failed. The Lesson Is Useful for Every Automaker.
Ford is not the only automaker that has stumbled with AI. GM tried replacing full-time engineers with contractors a few years ago to cut costs, and ended up with quality problems and an internal culture crisis before reversing course. FCA (now Stellantis) made a high-profile push toward automated assembly lines in the 2010s and saw assembly quality decline for three consecutive years. The pattern is consistent: companies underestimated how much tacit knowledge matters in manufacturing. The things that separate a well-assembled car from a problematic one are not written in any manual. They live in the experience of people who have spent years on the floor.
Charles Poon, VP of Vehicle Hardware Engineering, was more blunt than most executives get. The company had assumed that feeding design specifications into an AI system would reliably produce high-quality outputs. His assessment: “We were wrong about that.”
The outcomes have been positive. CEO Jim Farley said the returning gray-haired engineers are doing two jobs: training younger engineers and recalibrating the AI tools themselves. Warranty and recall costs have declined, contributing to what Farley described as hundreds of millions of dollars in cost improvements. Ford also moved up to first place among mainstream brands in the JD Power Initial Quality Study this year.
AI can accelerate many tasks in manufacturing. It cannot yet replicate the judgment that comes from decades of watching things go wrong and understanding why. Knowing which component will fail under specific conditions, which tolerance stackups create problems in humid environments, which sounds indicate a bearing going bad: these are things that require hands-on experience no current AI system has accumulated at scale. Gray-haired engineers are valuable because they can translate that experience into calibration signals for AI systems: telling the automated tools what a real problem looks like in sensor data, what variance is acceptable, what is not.
A significant portion of quality decisions in manufacturing are genuinely perceptual. The feel of a bolted joint. The sound a part makes under load. The way two surfaces meet. These are not abstract classification problems that get easier with more training data. They are physical experiences that humans process through sensory feedback in ways computer vision and sensors still cannot fully replicate. The engineers Ford called back have something that cannot be digitized and trained into a model: they have physically done the job and developed an intuition for when something feels right or wrong. That intuition, translated into calibration parameters for AI inspection systems, is more valuable than the inspection systems alone.
The right model is using AI to handle high-volume, repetitive tasks while keeping experienced humans in the loop for judgment-intensive quality decisions. Ford’s mistake was treating AI as a replacement for human expertise in a domain where judgment matters more than pattern matching. Poon comment about being wrong is a concise summary of what a lot of manufacturing companies are learning right now: AI is a powerful tool, but it is not a substitute for the accumulated knowledge of people who have been solving problems in a specific domain for decades.
For Chinese automakers aggressively deploying AI across their manufacturing operations, Ford’s experience is a useful reference. The wins from AI automation are real. So are the failure modes that come from assuming the technology can handle judgment it has not yet learned. The companies that get the balance right will have a real cost and quality advantage. The ones that treat AI as a people replacement rather than a people amplifier will end up calling the same kind of talent back, probably at higher salaries.
