Edition #8: They Bought the Judgment Back
The design requirements were in the system. The cars still came out wrong.
The Signal
Ford had the design requirements. Every one of them, written down, in the system. They fed them to the AI and waited for good cars to come out.
The cars came out wrong for three years.
Charles Poon runs vehicle hardware engineering at Ford. On a call with reporters on June 24, he said the company had mistakenly thought that introducing AI and ingesting the design requirements they had would produce a high-quality product. Then he said the part that matters. The AI was “only as good as the information you use to train it,” and in prior years Ford had not paid enough attention to the experience of its most knowledgeable engineers.
The information was there. The requirements were documented. What was missing had never been written down at all.
What was missing was knowing which tolerance looks fine on paper and cracks in the fourth winter, which supplier’s process drifts in August. None of that lives in a requirements document. It lives in someone who has been through enough product cycles to smell it before the numbers say anything.
So Ford went and bought it back. Over three years, 350 veteran engineers. Some were former Ford employees. Others came from suppliers, which is to say from the companies that build Ford’s parts. Ford calls them “gray beards” internally. They now work as internal auditors, running mandatory weekly design reviews to hunt for failure points before a blueprint reaches the plant floor.
It worked. In June, Ford came first among mainstream brands in the JD Power Initial Quality Study for the first time since 2010. Jim Farley told Bloomberg TV the effect on warranty and recall costs adds up to “hundreds and hundreds of millions” of tailwind on cost.
Here is the part nobody is pricing.
Ford did not train those 350 people. It bought them, on a market where they still exist, at a moment when they still exist. Some came off the retirement pile. Others came out of the supply chain, and those suppliers are now building Ford’s parts with that judgment gone from their own floor.
Meanwhile the machine that produces gray beards is being switched off. A Harvard working paper covering 66 million workers across more than 280,000 US firms found entry-level hiring at companies adopting generative AI fell roughly 80% per quarter since 2023, while senior headcount at the same firms kept growing. PwC’s 2026 AI Jobs Barometer found that in the most AI-exposed occupations, 52% of the new skills appearing in junior job ads were skills that used to be asked of experienced hires. In the least exposed, 7%.
Knowledge can be written down, and Ford proved it had been. Judgment gets built by doing the small work badly until you stop doing it badly, and the small work is exactly what is being automated first.
I watched this in the cloud wave. Data center sysadmins renamed themselves cloud engineers, and the workflows did look close enough. Up to a point they were. The concepts carried. Scale did not. The one worth having on your side was the one who had already scaled a platform and eaten the pain of it, and that never showed up on a CV. It showed up the first time something broke in a way the documentation did not cover.
Ford’s escape hatch worked because in 2026 there is still a stock of people who did that work, somewhere, that you can go buy. That stock was made in the 1990s. Nobody is making the 2040 batch.
Every board planning to automate now and hire the veterans back later is counting on the same shelf. It is not being restocked.
The Application
Ford. Over three years it hired 350 veteran engineers, a mix of former employees and people from its suppliers, after AI and automated quality systems failed to deliver. COO Kumar Galhotra said Ford had been relying more and more on automated quality systems and not getting the results it wanted. The veterans now hunt for failure points before parts reach the plant floor. Result: first among mainstream brands in the 2026 JD Power Initial Quality Study, 152 problems per 100 vehicles, 41 fewer than the prior year, first time on top since 2010. Ford has not published how many of the 350 came from suppliers, and no one has asked.
Emma Wiles, Boston University, with three researchers from BCG. Putting AI on the Org Chart. A survey of 1,261 managers found 23% already have AI agents formally listed on their org charts. In the experiment, managers got five documents with planted errors and 20 minutes to review them. The only thing that changed was the label: work from an AI tool, from an AI employee named ALEX-3, or from a human employee named Alex. Among managers whose companies had already put agents on the org chart, the AI employee label produced 18% fewer errors caught, 44% more requests for someone else to review, and a shift of perceived accountability away from the manager and toward the system. Among managers at companies without agents on the org chart, almost no effect. The authors are explicit that this is a governance decision.
Harvard, Hosseini and Lichtinger. Résumé and job posting data on 66 million workers across more than 280,000 US firms, 2015 to 2025. At firms adopting generative AI, entry-level hiring fell roughly 80% per quarter since 2023. Senior employment at those same firms kept rising.
The Noise
“Ford proves AI can’t replace people.”
This is the version that lets a room full of executives exhale, and it is the expensive one.
Ford reversed nothing. The AI quality systems are still installed and the investment continues. The gray beards were brought in to reprogram those systems and mentor junior staff, which means the veterans are being paid to encode the judgment that made them expensive. Ford is still the most recalled automaker in the United States, with 51 recalls this year covering more than 11 million vehicles, more than double the next manufacturer. Galhotra calls recalls a lagging indicator, which may be true and is also what you say when the number is bad.
Then there is the calendar.
On June 27, 2025, at the Aspen Ideas Festival, interviewed on stage by Walter Isaacson, Jim Farley said artificial intelligence was going to “replace literally half of all white-collar workers” in the US. Charles Poon made his confession on June 24, 2026. One year apart, minus three days, same company.
And by the time Farley said it, Ford had already been buying its engineers back for two years. The 350 were hired over three years, which puts the start around 2023. The CEO was forecasting the replacement of white-collar workers in public while his own operation was quietly paying to undo it.
The people his AI could not replace were engineers. White-collar workers. Ford ran the experiment on the exact population Farley said was finished, and had to buy them back at market price.
What Ford says is that its AI could not do this work without the people who knew what right looks like, and that it had let those people go before finding out.
The Question
Take the last thing you automated that is now running in production. Ask who signed off that the output was correct, by name. Who read it and put their name on it.
If the answer is the system, or if nobody can remember, the judgment left your company already and the invoice has not arrived yet. Ford’s arrived three years late and cost hundreds of millions.
Now What?
Write down the failure modes before you touch the process. Put the people who know what breaks in a room and make them document what they look for, in their words, including the things that sound like superstition. Ford paid for doing this in the wrong order. The window closes when they retire, and you do not get a warning.
Take the agents off the org chart. Wiles found the oversight damage only shows up where agents are formally institutionalized as employees. Reversing that costs you nothing but a naming convention. Call them tools, because they are tools, and keep the accountability on the person who used them.
Require a human name on every review of automated output. One person, signed. If nobody will put their name on it, nobody is reading it.
Look at your suppliers. Ford filled part of its gap out of its own supply chain. Whatever judgment you can buy there, your competitor can buy too, and your suppliers are running the same automation playbook you are. Ask them who reviews their output.
Count how many juniors joined your area in the last 24 months. That number is your stock of judgment in 2036. If it is zero, you have made a decision about 2036 without discussing it.
What I’m Watching
Robotics. On July 8, Nature published a UC San Diego preclinical trial where two teleoperated humanoids completed real gallbladder surgeries on large mammals, both in a human-robot pairing and robot-to-robot. Against a dedicated four-arm surgical system, a general-purpose humanoid is versatile, takes less room, and costs a fraction. The question stops being which robot does this surgery and becomes who can operate, and from where.
Health. On July 1 the FDA approved Casgevy for patients from age two with sickle cell disease or transfusion-dependent beta thalassemia, the first gene therapy cleared for children that young in these conditions. Roughly 5,500 newly eligible patients in the US across more than 75 authorized centers. CRISPR went from its first approval in late 2023 to early pediatric populations in under three years.
Quantum. The US signed two executive orders on June 22. One pushes commercialization and orders quantum sensors deployed by 2028. The other sets 2031 as the deadline for post-quantum authentication migration. Putting a date on it means the government thinks there is a real chance the hardware exists around then. Microsoft moved its own migration up to 2029, joining Google and Cloudflare.
This is my newsletter. Emerging technology without the hype. Real signals for strategic decisions.
Javier D’Ovidio