Charles Poon, Ford's vice president of vehicle hardware engineering, gave the plainest explanation for what went wrong that you're likely to hear from any large manufacturer this year: "Mistakenly, we thought that by just introducing artificial intelligence and ingesting the design requirements that we had, that would produce a high-quality product," he told reporters, as reported by Bloomberg. Ford had spent the past three years hiring back around 350 veteran engineers - staff internally nicknamed "gray beards" - after the AI-led quality control approach it had adopted failed to catch defects the way experienced human reviewers once did. TechCrunch's reporting and a Forbes analysis both land on the same detail: this wasn't a failed pilot quietly shelved. It ran long enough, and caused enough downstream cost, that reversing it became a multi-year hiring programme.
The problem wasn't the model. It was what left the building first
The sequencing is the part worth sitting with. Ford's experienced engineers left, through retirement and attrition, before their judgement had been captured anywhere the AI systems could draw on it. Design requirements got ingested as documents and specifications; the tacit knowledge behind those documents - why a particular tolerance existed, which supplier historically cut corners on a specific weld, what a subtle deviation in a part actually meant in practice - left with the people who held it. The AI tools didn't fail because they were bad at pattern matching. They failed because the patterns worth matching against had never been encoded, and a model ingesting a specification document has no way to reconstruct the judgement calls that specification was quietly resting on.
That produced exactly the failure mode you'd predict once you frame it that way: automated tools amplified weak inputs rather than catching design flaws, because nothing in the pipeline was positioned to catch problems the specifications themselves didn't anticipate. The rehired engineers now spend their time doing two things - catching defects before parts reach the plant, and retraining the AI systems and younger staff so the same knowledge gap doesn't reopen the next time someone retires. The result, according to CBT News, was Ford topping JD Power's 2026 Initial Quality Study in the mainstream category for the first time in 16 years.
Why this generalises well past manufacturing
Every enterprise running an AI programme built on "we documented the process, now the model can do it" is closer to Ford's original mistake than most would like to admit. Documentation is a compressed, lossy summary of expertise, written by people who assume the reader already shares their context. Feeding that compression into a model and expecting it to reconstruct the full judgement behind it is the same wager Ford made with vehicle design specs - and the same wager plenty of technology, finance and operations functions are quietly making with code review, underwriting, compliance sign-off and clinical or safety documentation right now.
The harder discipline Ford's experience points to is sequencing: don't let the people who hold undocumented judgement leave before you've built a genuine feedback loop between their review decisions and the system meant to eventually take some of that work on. A model trained purely on the artefacts a process produces - specs, tickets, requirements - is not the same as a model trained alongside the people making the judgement calls those artefacts don't fully capture. That loop needs to run for a meaningful stretch, with the expert still in the room able to say "no, that's wrong, and here's the part of this that isn't written down anywhere," before headcount reductions follow.
- Before reducing headcount on the strength of an AI tool's early performance, check whether the tool has actually been validated against the judgement calls your most experienced staff make that never made it into written procedure.
- Build a structured feedback loop where experienced reviewers correct the AI system's outputs for a meaningful period before their capacity is reallocated or reduced, rather than assuming ingesting documentation is equivalent to capturing expertise.
- Track quality or error metrics through any AI-assisted process transition specifically, rather than trusting that stable topline numbers mean nothing has been lost - problems introduced by a knowledge gap often surface downstream, not immediately.
- Identify which of your teams are about to lose institutional knowledge to retirement or attrition in the next 12-24 months, and treat capturing that knowledge as a dependency for any AI programme touching their function, not a parallel nice-to-have.
- Where AI tools are already running quality, compliance or review functions with reduced human oversight, ask what specifically would need to go wrong for the gap to be caught - if the honest answer is "nothing would catch it," that's the gap to close first.
Ford's fix wasn't more AI, and it wasn't abandoning AI either - it was accepting that the automated tools needed the very expertise they were meant to reduce demand for, kept close enough to keep correcting them. Want a candid view on whether an AI-assisted process in your organisation still has the expert feedback loop it needs, or lost it somewhere along the way? Email sales@halfteck.com.