Meta Project OT: The AI Workforce Experiment That Hit a Wall

Artificial intelligence is increasingly being positioned as more than a productivity tool. The newest generation of AI agents is designed to perform tasks autonomously—writing code, analyzing information, operating software, handling customer interactions and potentially replacing entire workflows.

Few companies have embraced that vision more aggressively than Meta.

In early 2026, Mark Zuckerberg and Meta executives developed an ambitious internal initiative called Project OT, short for Organization Transformation. The objective was to redesign parts of Meta around AI, reduce the size of traditional teams, remove layers of management and allow small groups of employees to accomplish substantially more with AI agents.

The idea sounded simple:

Fewer people + smaller teams + AI agents = dramatically higher productivity.

But the implementation ran into serious problems.

Internal data reviewed by Reuters indicated that AI-assisted code production increased dramatically, while the increase in actual user-facing features was considerably smaller. Reliability and security problems also increased, while employee morale deteriorated.

The result was not the AI-native workplace Meta had hoped for.

Instead, Zuckerberg ultimately scaled back the most aggressive parts of the plan and canceled a planned second wave of restructuring.

What Was Project OT?

Project OT was Meta’s internal Organization Transformation initiative.

It was designed around the idea that AI could fundamentally change how a large technology company operates.

Rather than maintaining traditional teams containing engineers, product managers, designers, data scientists and other specialists, Meta experimented with much smaller AI-native pods.

A traditional product-development team might contain roughly 10–20 people with specialized roles.

The proposed AI-native model could instead involve approximately 3–5 people, supported by AI tools and shared specialists.

The underlying philosophy was:

AI should allow a small number of highly capable employees to accomplish work previously requiring much larger teams.

Reuters reported that some internal scenarios considered reducing the size of certain teams by as much as 60%.

Importantly, this did not mean Meta planned to fire 60% of its entire workforce. Meta explicitly said the scenarios applied to particular teams and included a combination of layoffs, redeployments and unfilled positions.

Why Was Meta Betting So Heavily on AI?

The timing wasn’t accidental.

Meta was simultaneously committing enormous amounts of money to artificial intelligence infrastructure.

The company expected to spend at least $130 billion on AI infrastructure and chips in 2026, according to Reuters reporting.

That created pressure to demonstrate that such massive AI investment could produce corresponding productivity improvements.

The logic was straightforward:

If AI infrastructure costs billions, the organization using that infrastructure must also become substantially more productive.

Project OT was therefore not simply an experiment with chatbots.

It was an attempt to redesign how a major technology company actually worked.

The Hawaii Leadership Retreat

The Project OT strategy began taking shape during Meta’s leadership retreat in Hawaii in January 2026.

Zuckerberg and senior executives discussed how Meta could become an “AI-native” company.

The vision involved:

• Smaller teams
• Fewer management layers
• AI-assisted development
• Autonomous AI agents
• More fluid employee roles
• Greater use of AI in daily workflows
• Increased automation
• More output from fewer employees

The ultimate ambition was much larger than simply giving employees access to ChatGPT-style tools.

Meta wanted AI to become part of the organization’s operating structure.

From Traditional Teams to AI-Native Pods

One of the most interesting elements of Project OT was the proposed pod structure.

Traditional product development generally depends on clearly defined roles:

• Engineers
• Product managers
• Designers
• Data scientists
• Researchers
• Data engineers

The AI-native model attempted to blur those boundaries.

A small group of “builders” could use AI systems to perform tasks that previously required several specialized employees.

Shared specialists could support multiple pods rather than belonging permanently to a single team.

The organizational chart would therefore become flatter.

Instead of large teams with multiple layers of management, Meta envisioned numerous small pods reporting upward to a limited number of organizational leaders.

Reuters reported that at least 11 Meta units had implemented versions of these smaller pod structures by June.

The “10X Employee” Concept

Another important part of the transformation was the search for exceptionally productive employees.

Meta’s HR organization developed a concept around identifying “Irreplaceable Talent.”

The idea was that AI could potentially turn an exceptionally capable employee into something resembling a 10X worker.

In theory:

1 highly capable employee + AI = output of a much larger team

That model has obvious financial appeal.

Instead of paying 10 people to perform a workflow, a company could potentially pay one exceptional employee equipped with powerful AI systems.

But there is a fundamental problem.

Human productivity isn’t simply a function of how much code or content an employee can generate.

Quality, coordination, judgment, security, testing, product-market fit and accountability also matter.

Project OT exposed that distinction.

The Layoffs

Meta eventually proceeded with a significant workforce reduction.

On May 20, 2026, the company cut approximately 10% of its workforce.

However, the more aggressive second wave of restructuring planned for later in the year was canceled shortly before the first layoffs occurred.

According to Reuters, Zuckerberg made the decision to abandon the planned November wave shortly before the May layoffs.

This distinction matters.

The story isn’t:

“Meta fired 8,000 people and completely abandoned AI.”

The more accurate story is:

Meta implemented a major workforce reduction while simultaneously discovering that its most aggressive AI-driven restructuring assumptions weren’t working as expected.

Employees Were Also Being Asked to Train AI

One particularly controversial element involved employee activity data.

Reuters reported that Meta had mandated tracking software on some U.S. employees’ devices to capture information such as keystrokes and mouse movements for AI training.

The goal was to help AI agents learn how humans interact with computers.

Employees understandably interpreted this as potentially helping train systems that could eventually automate their own work.

That created a serious trust problem.

Instead of thinking:

“AI will help me do my job.”

Some employees began thinking:

“I’m helping train the system that will replace me.”

That distinction is critical when implementing workplace AI.

Employee Morale Collapsed

The restructuring produced significant internal backlash.

Meta’s internal employee sentiment score reportedly fell from 74% favorable to 55% favorable.

Employees were frustrated by:

• Uncertainty surrounding layoffs
• Communication problems
• AI-driven restructuring
• Tracking concerns
• Job reassignment
• New organizational structures
• Anxiety about AI replacing human workers

The technological transformation had therefore created an organizational problem of its own.

AI wasn’t just changing workflows.

It was changing how employees perceived their relationship with the company.

Then Came the Productivity Problem

This is arguably the most important part of the entire story.

AI was producing more code.

Much more code.

According to internal information reported by Reuters, changes to Meta’s internal software platforms and infrastructure increased by approximately 220% year over year.

At first glance, that sounds extraordinary.

But another metric told a very different story.

Changes that actually resulted in new or upgraded features reaching Meta users increased by only 36%.

That creates a massive gap:

Code generated: +220%

User-facing feature growth: +36%

The lesson is important.

More Code Does Not Equal More Productivity

AI makes it incredibly easy to generate code.

But software development isn’t simply about writing code.

A real production system also requires:

• Architecture
• Testing
• Code review
• Security
• Monitoring
• Debugging
• Deployment
• Documentation
• Reliability engineering
• Product decisions
• User feedback

If AI increases the amount of code being produced without increasing useful product output proportionally, the organization may simply be generating more work to manage.

That appears to have been one of the central problems inside Meta.

AI Agents Created Reliability and Security Problems

The problems weren’t limited to productivity.

Reuters reported that Meta’s internal posts described reliability warning signs associated with the increase in AI-generated code.

There were also warnings about AI agents performing disruptive actions at a scale that humans were unlikely to execute.

Major technical and security incidents reportedly increased by approximately 40%, while the amount of employee time spent dealing with those incidents increased by about 70%.

That creates another equation:

More AI automation → more automated activity → more unexpected behavior → more human firefighting.

When that happens, automation stops being a productivity multiplier.

It becomes another operational burden.

The Instagram Security Incident

The problem became more visible in June.

Reuters reported that hackers exploited Meta’s new AI-powered customer-support chatbot to gain access to high-profile Instagram accounts, including the dormant official White House account associated with the Obama administration.

The incident demonstrated one of the biggest risks of autonomous AI systems:

An AI system that has permission to act can potentially create much greater damage than an AI system that only provides information.

A chatbot that answers a question is one thing.

An AI agent capable of taking actions across systems is fundamentally different.

The more permissions an agent has, the more important security controls, authentication, monitoring and human oversight become.

Zuckerberg Eventually Admitted the Technology Wasn’t Moving Fast Enough

By July, Zuckerberg acknowledged that the technology had not advanced as quickly as he had expected.

He said AI-agent technology had not “accelerated” at the pace he had anticipated and suggested the technology could improve substantially over the following three to six months.

This wasn’t a rejection of AI.

It was a recognition that AI capability and organizational readiness were not progressing at the same speed.

Meta Didn’t Abandon AI

This distinction is extremely important.

Project OT’s setbacks do not mean Meta has stopped investing in AI.

Quite the opposite.

Meta continues to make enormous investments in AI infrastructure, models and AI agents.

In September 2026, Meta also launched Muse, an AI assistant capable of autonomously performing tasks across connected applications, including areas such as email, travel, payments and shopping. Reuters reported that internal testing had uncovered reliability and security issues before the product’s launch.

So the real lesson isn’t:

“AI doesn’t work.”

The lesson is:

“AI agents aren’t yet a frictionless substitute for complex human organizations.”

Why the Experiment Matters

Meta’s experience provides a useful case study for almost every company considering aggressive AI adoption.

AI can dramatically increase the speed at which people produce things.

But production volume isn’t the same as business value.

Consider a simple example.

Imagine a company previously had:

10 developers → 100 units of useful output

AI could potentially allow those developers to generate:

300 units of code

But if only:

120 units become reliable, secure, valuable products

then the organization hasn’t achieved a 3X productivity improvement.

It may simply have created a larger pipeline of unfinished or risky work.

The AI Productivity Paradox

This creates what we can call the AI productivity paradox:

AI can increase activity faster than it increases value.

Employees can:

• Write more code
• Generate more documents
• Create more prototypes
• Produce more designs
• Analyze more data
• Generate more marketing assets

But organizations still need to answer:

Did customers receive more value?

That’s the metric that matters.

AI Should Augment Humans Before It Replaces Them

For most businesses, the better strategy is likely to be augmentation rather than immediate replacement.

Instead of:

AI replaces employee

a more sustainable model is:

Employee + AI + automation = higher-capacity employee

For example:

Customer service

Instead of replacing the receptionist entirely:

AI can answer routine questions, collect information and route calls while the human handles complex cases.

Marketing

AI can generate initial campaign concepts while marketers handle positioning, brand strategy and approval.

Software

AI can generate code while engineers remain responsible for architecture, security, testing and production deployment.

Accounting

AI can automate document extraction and reconciliation while accountants handle judgment-heavy compliance and advisory work.

The objective should be to eliminate low-value repetitive work, not blindly eliminate humans.

What Businesses Should Learn From Meta

1. Don’t Measure AI by Activity

Don’t ask:

“How much content did AI generate?”

Ask:

“How much valuable output did AI create?”

2. Keep Humans Responsible for High-Risk Decisions

AI can recommend.

Humans should remain accountable for decisions involving:

• Security
• Finance
• Legal compliance
• Hiring
• Customer disputes
• Production systems
• Sensitive information

3. Don’t Give AI Unlimited Permissions

Agentic AI becomes dramatically more powerful when it can take actions.

But permissions should be carefully controlled.

Use:

• Least-privilege access
• Approval workflows
• Logging
• Sandboxing
• Monitoring
• Human escalation
• Automated rollback

4. Measure Reliability Alongside Productivity

If AI increases output but also increases incidents, the net productivity gain may be negative.

Track both:

Output ↑

and

Errors ↓

5. Don’t Confuse AI Adoption With AI Transformation

Installing an AI tool doesn’t make a company AI-native.

The real transformation involves:

People + processes + technology + governance + culture.

Ignoring any one of those components can create serious problems.

The Bigger Question: Will AI Replace Jobs?

The Meta story doesn’t prove that AI will never replace jobs.

It demonstrates something more nuanced.

AI can automate tasks much faster than it can reliably replace entire human roles.

A job is usually a collection of different tasks.

Some can be automated easily.

Others require:

• Judgment
• Context
• Communication
• Creativity
• Accountability
• Negotiation
• Human trust

The future is therefore likely to involve substantial job redesign, even where complete job replacement doesn’t happen.

Meta’s Experiment Is Not Over

Calling Project OT a complete failure would also be too simplistic.

Reuters’ September follow-up noted that the initiative may be better described as being scaled back or put on hold, rather than permanently dead. Meta continues to experiment with AI-native teams and autonomous systems.

The company still believes AI can fundamentally change how work gets done.

The difference is that the most aggressive assumptions have been tested against reality.

And reality turned out to be considerably messier.

The Real Lesson From Zuckerberg’s AI Experiment

The biggest lesson isn’t:

“Humans beat AI.”

And it isn’t:

“AI will replace everyone.”

The real lesson is:

AI is powerful enough to transform work—but not yet predictable enough to remove humans from every critical part of the system.

Meta had enormous capital, world-class engineers and access to cutting-edge AI technology.

Yet the company still encountered:

• Productivity gaps
• Reliability problems
• Security incidents
• Employee resistance
• Communication failures
• Organizational confusion
• Unexpected operational costs

That should make smaller companies cautious about assuming that simply adding AI agents will automatically reduce headcount and costs.

What Comes Next?

The next stage of AI adoption is likely to be less about “replace everyone” and more about “redesign the workflow.”

The winning companies may not be those that fire the most employees.

They may be those that figure out how to give each employee significantly more leverage.

The future could therefore look less like:

Human vs AI

and more like:

Human + AI vs Human without AI.

That is a much more useful way to think about the coming AI transformation.

Final Takeaway

Meta’s Project OT is one of the clearest real-world experiments in what happens when a major technology company attempts to move rapidly toward an AI-native organizational model.

The experiment produced some impressive gains in AI-assisted activity.

But it also revealed a critical limitation:

Generating more output is not the same as generating more value.

Meta’s experience suggests that the immediate future of AI at work may not be a world where humans disappear.

Instead, it may be a world where small teams equipped with powerful AI systems become dramatically more capable—while humans remain responsible for judgment, oversight, security and accountability.

And that may ultimately be the more important AI revolution.