Why Meta's AI Layoffs Backfired: The Judgment Problem

Meta Wrote 220% More Code Last Year. It Shipped 36% Less. Now You Know Why the Layoffs Stopped.

Meta wrote 220% more code in 2024 than 2023. It shipped 36% more features.

Sit with that gap for a second.

Between writing and shipping sits the entire story of why Meta's AI-first restructuring — code-named Project OT, leaked to Reuters last month — hit the emergency brake after the first round of cuts. Why incidents spiked 40%. Why firefighting jumped 70%. Why Zuckerberg pulled the plug on wave two hours before it hit.

Everyone read it as "the AI wasn't ready yet."

That's not what happened.

The bottleneck didn't disappear when AI showed up. It just moved. And Meta fired the people standing where it moved to.

I've watched this movie four times now. The railroad version. The digital photography version. The algorithmic trading version. Every time, we mistake the capability for doing the work with the capability for knowing which work is right. Every time, we learn the difference too late.

The Plan Nobody Questioned (Until It Failed)

Project OT was elegant on paper. Make Meta "AI native." Collapse engineer, designer, and data scientist into one role — "builder" — and drop them into small pods. Each pod reports to one leader managing fifty people. Cut middle management by 60% in some divisions. Two waves of layoffs, six weeks apart.

The logic was clean: if AI agents can write code, review pull requests, generate designs, and run data analysis, why do we need three separate roles? Why do we need managers between the work and the decision?

Meta had conflated "producing output" with "producing value." The AI could write the code. It could not tell you which code to write. Or which of the seventeen implementations it generated would fail in production. Or which feature request was solving the wrong problem.

That gap — between output and judgment — used to live in two places. Middle management held some of it. The entry-level work held the rest.

The Rung You Can't See Until It's Gone

Here's what Meta discovered in real time: the junior work wasn't waste. It was the training ground for judgment.

The reconciliation nobody wanted to do. The bug that took three days to isolate. The customer complaint that didn't fit the feature spec. The pull request review that caught the edge case. That's where you learned to tell good code from code that compiles.

Cut the bottom rung of the ladder and suddenly there's no way to reach the middle. You're asking AI to flood the system with options and hoping someone ten years into their career can catch what's wrong — except you just fired half the people who spent a decade learning how.

I was on a call last week with a financial services client trying to figure out why their AI-generated reconciliation reports were creating more work than they saved. The answer was in the room: they'd moved two junior accountants off recs and onto "AI oversight," but neither had done a manual rec in eighteen months. They could spot a formatting error. They couldn't spot a methodological one.

The senior person who could? She was spending 60% of her week reviewing AI output instead of the 20% she used to spend reviewing junior staff work. The AI produced more. It required deeper review. And there were fewer people capable of doing it.

Where Judgment Lives (And Why You Can't Buy It)

Meta has $130 billion in cash. It tried to buy a solution by hiring "senior builders" to replace the middle managers. It didn't work.

Judgment isn't a hiring problem. It's a growing problem.

You can't parachute someone in to judge work they didn't learn to do. The person who can review an AI-generated financial model learned it by building fifty financial models badly first. The person who can catch the subtle error in a code review spent two years writing code someone else caught errors in.

Middle management wasn't a cost center. It was where the company stored judgment. The manager who could look at three competing implementations and say "this one will break under load in six months" learned that by shipping the one that broke under load. You can't replace that with a policy document or a senior hire.

Zuckerberg called off wave two because incidents spiked and the people left couldn't keep up. But the real problem was structural: he'd removed the system that created the people who could keep up.

The Question You're Not Asking

Your firm is looking at AI and asking "what work can this do?"

That's the wrong question.

The question is: who's left who can tell when it's wrong?

Not "can AI draft the audit summary?" but "who in your org can spot the material omission in the AI-drafted audit summary — and do you still have the pipeline that creates that person?"

Not "can AI generate the financial model?" but "who reviews the model, and what happens to your bench when they retire?"

Not "can we cut headcount?" but "are we cutting the training ground for the judgment we'll need in three years?"

I've seen this pattern before. In 2008, firms automated trade execution and cut junior trading staff. By 2015, they had a crisis: nobody left who understood market structure well enough to diagnose the flash crash. The people who could see it had learned by doing it manually. The people hired after automation went straight to oversight without building the underlying map.

What Actually Works (And What Doesn't)

Here's what I'm telling clients right now:

Don't automate the work and cut the people. Automate the work and change what the people do. The junior accountant stops doing reconciliations and starts reviewing AI-generated recs — but they're still learning what a reconciliation is, still building the pattern recognition that makes them useful in year five.

Track judgment, not output. If your AI is producing 200% more and your team is shipping 30% more, someone is doing 170% more review. Name that work. Staff it. Treat it like the asset it is.

Map your bench. Who in your firm can catch a material error in [the work AI is doing]? How many? What's their average age? If they left tomorrow, how long to replace them? If the answer is "we'd struggle," you have a judgment gap, not a hiring gap.

Protect the training ground. The boring work teaches pattern recognition. If you automate it away, create something else that builds the same skill. Residencies. Rotations. Supervised reviews with feedback loops. Something.

Nobody gets fired the day the AI arrives. The gap just grows quietly until something breaks and there's nobody left who knows how to fix it.

What to Do Monday Morning

Walk into your next AI implementation meeting and ask:

• What work is this replacing?

• Who learned their judgment by doing that work?

• Where will we grow that judgment now?

• Who's left who can review this output — and what happens when they retire?

If your AI strategy doesn't answer those four questions, you're not implementing AI. You're building the gap Meta just discovered.

I've been through railroad disruptions and newspaper disruptions and trading floor disruptions. The technology always works eventually. The question is whether you still have the people who know what "works" means when it does.

Meta can afford to learn this lesson in real time. Can you?