Ford Didn’t Have an AI Problem. It Had an Information Problem.

It Had an Information Problem.

Ford just told on itself, and the admission is worth sitting with.

The company recently disclosed that it has rehired more than 300 veteran quality inspectors in recent years, walking back a bet it made on AI-driven quality checks. The explanation from Charles Poon, the company’s vice president of vehicle hardware engineering, is the part that matters: “Artificial intelligence is a fantastic tool, but it’s only as good as the information you use to train it. Over prior years, we didn’t pay as much attention as we should have to the experience of our most knowledgeable engineers that have been with us through many product cycles.”

Translate that out of corporate-speak. Ford automated a function before it had captured the knowledge the function depended on. The veteran inspectors who knew what a quality defect actually looked like, the kind of judgment built over decades of product cycles, weren’t systematized before the people who held it walked out the door. By the time the company tried to train its AI tools properly, some of that knowledge was simply gone.

This is not a story about AI failing. It’s a story about an organization automating a decision before it had built the infrastructure to make that decision portable.

The pyramid problem

I’ve previously referenced the Gartner AI-human interaction pyramid: Tool, Assistant, Collaborator, Companion. Every AI deployment decision is implicitly a bet about where on that pyramid a given function actually sits. Ford’s bet was that quality inspection could function near the Tool end, that ingesting design requirements would be sufficient to produce a high-quality product.

That bet failed because quality inspection, it turned out, wasn’t really a Tool-tier function. It depended on tacit judgment, the kind that experienced engineers carry but rarely write down, because nobody ever built the system that would have asked them to. The AI wasn’t undertrained because the technology was immature. It was undertrained because the organization had no mechanism for moving expert judgment out of people’s heads and into a form a system, or a successor, or a process could actually use.

That is an infrastructure failure, not a technology failure. And it is the same failure I’ve described in other contexts: information existed inside the organization, but the organization had no infrastructure for it to travel to where decisions were being made.

Centralize standards, decentralize information

This case also sharpens a principle I’ve been developing: centralize standards, decentralize information. Ford centralized the standard, the design requirements, and assumed that was the hard part. It wasn’t. The harder problem was capturing the decentralized information, the on-the-floor judgment of thousands of individual inspections across product cycles, and feeding it back into the system that was supposed to replace those inspectors.

You don’t get a high-quality AI quality-checker by handing it the spec sheet. You get one by building the infrastructure that lets the people closest to the work transmit what they know before they leave. Ford’s 900 AI-powered cameras could detect anomalies. They couldn’t tell you, the way a twenty-year veteran could, which anomalies actually mattered.

The correction is the real story

What makes this case study useful rather than just a cautionary tale is what Ford did next. The company didn’t abandon AI. It rehired the veteran engineers specifically to train its systems and mentor younger workers, treating institutional knowledge as an input to build rather than a cost to cut. Ford also reported reclaiming the top spot in the JD Power Initial Quality Study for the first time since 2010, attributing it in part to what it called a “significant talent refresh.”

That’s the organizational design lesson. The fix wasn’t less AI. It was building the missing layer, the one that moves expert judgment into a form the rest of the system can use, before asking automation to carry weight that infrastructure was never built to support.

Every organization currently deploying AI is making the same implicit bet Ford made: that a given function sits further up the Tool-to-Companion pyramid than it actually does. Ford found out the expensive way what that bet costs when it’s wrong.

Companies are racing to automate work. Far fewer are taking the time to understand what their experts are actually contributing before replacing pieces of it. If you don’t understand where judgment lives, you won’t know you’ve removed it until something fails. By then, the expert you eliminated has become the consultant you hire back.

Technology will continue improving. But organizations that outperform won’t be the ones that replace people the fastest. They’ll be the ones that first identify where experience creates value, preserve that advantage, and then use AI to amplify it instead of assuming it can substitute for it.