Meta Halts Second Round of Layoffs: AI Employees Fall Short of Expectations
nashnova research
Meta has shelved a second round of layoffs planned for November after its internal AI agents failed to deliver; code changes surged 220%, but user-facing new features grew just 36% — the company's own data is undermining the case for replacing humans with AI.
Why did Meta hit the brakes on layoffs?
In January, Zuckerberg and executives discussed "Project OT" (Organizational Transformation) at a Hawaii retreat. The plan: shrink 10–20-person product teams into 3–5-person "Pods," with some teams facing cuts of up to 60%.
The premise was that AI agents could absorb the workload. In April, Meta required employees to install tracking software that logged keystrokes and mouse clicks — data to train AI agents to mimic how humans operate computers.
Zuckerberg admitted at an internal meeting: the timing of the restructuring was misjudged, and agent technology advanced slower than expected. The layoff plan was frozen.
Code volume up 220% — what did users actually get?
Internal data disclosed by Meta's CTO: after scaling up AI, code changes rose 220% year-on-year, but changes that delivered user-visible new features rose only 36%.
This means → most AI-generated code never turned into product value — it piled up in infrastructure.
Worse still, major technical and security incidents rose 40% year-on-year, and time spent on incident response rose 70%. In plain terms = AI wrote more code but also created more problems, leaving people busier than before.
What emergency corrections did Meta make?
Alongside halting layoffs, Meta paused the mouse-tracking project, sharply cut spending on Anthropic, and allowed some employees reassigned to AI engineering teams to return to their original roles.
The company improved office perks and increased travel and social-event budgets to steady morale. This reflects an internal red flag — the semi-annual survey showed positive sentiment falling from 74% to 55%.
Is Microsoft hitting the same wall?
Microsoft's Chief Strategy and Transformation Officer admitted the company deployed AI tools the same way it rolls out traditional software — deploy, train, promote — but AI usage quickly plateaued, and the expected business results never materialized.
LinkedIn CEO Ryan Roslansky described a telling scenario: he received an AI-generated document from a colleague, then sent it to his own AI for a summary. "Back and forth, zero new ideas were produced." He called it a "death spiral."
In plain terms = AI writing for AI, AI summarizing for AI — information spinning in a loop with no one actually thinking.
Where did the money go?
Microsoft executives initially estimated internal spending on Anthropic technology could reach at least $1 billion per year. That estimate was later cut by more than a third.
Some employees previously had AI model budgets of up to $100,000 per person per month, now reduced to roughly $10,000. This means → per-person AI spending shrank by 90%, a sign that earlier usage was extremely wasteful.
Microsoft is now steering employees toward its own GitHub Copilot and requiring departments to manage their token budgets.
What does this mean for the broader enterprise AI market?
An EY survey published in September found that among 202 senior AI decision-makers at large U.S. public companies, 36% said their organization had experienced an AI incident causing material negative impact — including data loss, financial losses, operational disruptions, and reputational damage.
SHRM CEO Taylor compared the current state of enterprise AI deployment to "driving a truck down the highway while still building the truck."
This reflects a pattern beyond Meta and Microsoft — whether the AI-efficiency narrative can actually deliver inside enterprises remains the biggest unproven thesis in tech right now.
市场有风险，内容仅供研究参考，不构成投资建议。