What Ford got wrong
Ford's quality problem was not that the AI tools were bad. It was that the AI tools were built without the knowledge the most experienced engineers carried. Many of those engineers had left Ford before their institutional knowledge could be captured or used to train the systems meant to replace them. The automated quality systems that followed were checking for what they had been told to check for. They were not checking for what the veterans had learned to look for over decades: the subtle indicators, the edge cases, the failure modes that do not appear in a written design requirement. Charles Poon was direct about what that required to fix: "We recognised that for us to enhance some of our automation and machine learning and artificial intelligence tools we needed to ensure that they were trained by the most experienced individuals." The gap between what the AI was checking and what experienced engineers would have caught showed up in the recall numbers, in warranty costs, and in J.D. Power's rankings, where Ford had been sliding for years before the rebuild began. The three-year correction required hiring back the people who had already left, using their knowledge to retrain the systems, and building what should have been built before the original automation was deployed.
Why this is not just a manufacturing story
The specific failure mode Ford experienced is not unique to vehicle assembly lines or quality inspection systems. Any process that depends on experienced judgment carries the same risk when AI is introduced without first capturing what that judgment involves. Consider a few patterns that appear regularly in smaller businesses. A business uses AI to handle initial customer support queries, but the experienced team member who used to handle them had developed a working sense of which situations required escalation and which did not. That pattern was never documented. The AI handles standard queries. Unusual situations fall through. A business automates invoice processing, but the accounts payable staff member who used to flag vendor inconsistencies and billing anomalies is no longer reviewing them. The AI processes everything that matches the expected format. A business uses AI to screen job applicants, but the experienced hiring manager who used to identify non-obvious candidates and flag certain patterns is no longer in the early review step. The AI filters against explicit criteria. In each case, the knowledge that made the experienced person effective did not transfer to the AI automatically. It requires explicit, deliberate transfer before the automation can approach the level of judgment the person it replaced was exercising.
What business owners often misunderstand about automating expertise
The first misunderstanding is that AI learns from watching. Most business AI tools do not learn from being deployed in your operation. They were trained on data before you set them up. If that training data does not include your specific processes, your edge cases, and your exception patterns, the AI has no way to develop the judgment the experienced person had. It can handle what it was trained on. Domain-specific knowledge is usually not in that training set unless you put it there. The second misunderstanding is that AI errors in expertise-dependent processes look like obvious failures. When an experienced engineer caught a defect, the catch was visible. When the AI missed the same defect, the miss was not visible until the vehicle reached the customer. Errors in processes that depend on experienced judgment can be hard to detect precisely because the person who would have caught them is no longer in the loop. The absence of a catch does not produce an alert. The third misunderstanding is that documentation can happen after the automation is running. Several businesses plan to document processes after they see how the AI performs in practice. Ford illustrated what that approach costs when the expert knowledge is already gone: the documentation had to be reconstructed by rehiring the same people, at ongoing cost, after quality had already declined and the cost of the errors had already accumulated.
Three questions to answer before automating any process that depends on experienced judgment
The first question: is the knowledge that drives this process documented in a form an AI tool can actually learn from? Not a general process manual. Not a job description. The specific edge cases, exception patterns, failure modes, and judgment calls that the experienced person handles routinely. If that knowledge is not documented, the AI operates without it. If the documentation does not exist yet, it needs to be created before any automation decision is finalized. The second question: who catches the AI's errors, and how? Before automating a process, define who reviews AI output, at what frequency, and with what level of domain knowledge. This review function requires someone who genuinely understands the process, not someone checking whether the AI produced output in the expected format. Errors in expertise-dependent processes are often noticeable only to someone who knows what correct looks like in cases that were not anticipated. The third question: what happens when the AI encounters a situation it was not trained for? If the experienced person is no longer available and there is no escalation path, there is no fallback. Design any automation so that situations outside the normal pattern are escalated to a human with domain knowledge, not handled silently by a system that was not trained for them.
What a serious business should do next
If your business is planning to automate a process that currently depends on experienced judgment, the right first step is documentation of that judgment before any tool selection begins. That documentation is not a process map. It is a record of what the most experienced person in that role does when a situation is not standard: the exception cases, the patterns they look for that a less experienced person would miss, the rules that exist only in their working knowledge and have never been written down. If the experienced person who holds that knowledge is still with the business, involve them in the AI setup process before any transition happens, not after. Their role in training the AI or reviewing its early output is not optional work. It is the mechanism by which the AI develops any meaningful competence with the domain-specific judgment the business depends on. If the experienced person has already left, the situation is more expensive to correct, as Ford found. The realistic options are to bring that knowledge back in some form, run the AI in a supervised mode while domain knowledge is rebuilt, or accept that the automation will not perform at the level of the person it replaced and scope the deployment accordingly. For businesses that are not sure which processes are strong candidates for AI automation and which require this level of preparation, a diagnostic review of the current situation is a useful first step before any tooling decision is made.
The Atlacis view
Ford's story is useful because the executives were direct about what went wrong. Most companies that make this mistake do not make public statements about it. The errors show up as cost overruns, quality drift, customer complaints, and turnover without anyone explicitly connecting them to the automation decision. The pattern runs the same way at any size: automate a process that depends on experienced judgment, lose the experts before the knowledge is captured, discover the gap when errors show up in ways that are hard to trace back to the automation decision, and spend more on the repair than the original automation was expected to save. The right order is different. Understand the process before selecting any tool. Document the judgment calls before the experts leave. Build the AI around the documented knowledge. Verify it handles the hard cases before removing the human review layer. Ford still uses AI in its quality operations today. The point is that the AI now works because it was trained on real expert knowledge rather than being deployed as a substitute for expertise that was never captured. At Atlacis, we help business owners understand which processes are strong candidates for AI automation and which require this kind of preparation first. If you are considering automating work that currently depends on experienced people, the right starting point is a clear look at what those people actually know and whether that knowledge has a documented path into the system before any transition begins.
The short version
- Ford VP Charles Poon admitted the company mistakenly believed AI would produce high-quality vehicles after being fed design requirements alone. It did not. Ford spent three years and billions of dollars in recall costs correcting that mistake.
- The core failure was not the AI tools. Experienced engineers left Ford before their institutional knowledge could be documented or transferred into the systems designed to replace them. The AI had no expert knowledge to draw from.
- The fix required rehiring around 350 veteran engineers, known internally as 'gray beard' specialists. Their job was to retrain the AI tools, run mandatory quality reviews, and rebuild the quality process around captured expert knowledge.
- The result: Ford topped J.D. Power's 2026 Initial Quality Study among mainstream brands for the first time in 16 years, improving from No. 15 in 2023 to No. 1 in 2026. Only Porsche and Genesis ranked higher overall.
- The lesson applies to any process that depends on experienced judgment. AI does not absorb expertise by being deployed in a role. The knowledge has to be explicitly documented and transferred before or during the AI setup process.
- Before automating any expertise-dependent process, answer three questions: is the knowledge documented in a form the AI can learn from; who catches errors the AI was not trained for; and what escalation path exists when the AI encounters an edge case.
Where ATLACIS can help
Sources
- Ford official blog: Ford Named Top Mainstream Brand in 2026 JD Power Initial Quality Study (June 25, 2026)
- J.D. Power 2026 U.S. Initial Quality Study press release (June 25, 2026)
- TT News (Bloomberg/Keith Naughton): Ford Has Been Rehiring Quality Inspectors as AI Falls Short (June 25, 2026)
- TechCrunch: Ford rehires 'gray beard' engineers after AI falls short (June 28, 2026)