Meta spent much of 2026 trying to answer one of the most consequential questions facing the technology industry: How many people does a major technology company really need when AI agents can perform an increasing share of knowledge work?
The answer appeared, for a time, to be dramatically fewer.
Internal documents reviewed in a Reuters investigation reveal that Meta developed Project OT, short for Organization Transformation, as an ambitious blueprint for turning the company into an “AI-native” organization. Some teams could potentially be reduced by as much as 60%. Traditional management layers could disappear. Product groups could shrink into tiny teams, or “pods,” supported by autonomous AI agents.
Then something unexpected happened.
The AI-generated work began increasing dramatically. But the value delivered to users did not increase at anything close to the same rate.
Internal figures reportedly showed code changes to Meta’s internal software platforms and infrastructure rising 220% year over year, while new or upgraded features reaching users increased only 36%. Major technical and security incidents reportedly increased 40%, while employee time spent dealing with those incidents rose as much as 70%.
That gap between activity and value became the central problem.
Hours before Meta implemented its first restructuring wave on May 20, CEO Mark Zuckerberg reportedly canceled a second wave that had been planned for November.
The first wave still went ahead.
Approximately 8,000 employees—around 10% of Meta’s workforce—lost their jobs, while roughly 7,000 others were moved into AI-related roles. Meta also canceled approximately 6,000 open positions.
The story is therefore not simply about layoffs.
It is a revealing case study in what happens when an organization tries to move from AI-assisted work to AI-directed work faster than its technology, operating model and workforce can adapt.
The ambition behind Meta Project OT AI workforce transformation
Project OT emerged from an internal effort to rethink how Meta should operate in an era of increasingly capable AI agents.
According to Reuters, the thinking developed following a January 2026 leadership retreat in Hawaii. The underlying proposition was straightforward but radical: if AI agents could perform substantial portions of engineering, analysis, product development and other knowledge work, Meta could operate with significantly smaller human teams.
The proposed organization looked very different from the conventional technology-company model.
Instead of large product organizations with multiple layers of managers, specialists and individual contributors, the internal “AI-Native Playbook” envisioned smaller groups in which humans would supervise AI systems capable of performing significant amounts of the actual work.
Some teams could operate with only three to five people.
Roles could become broader. The distinction between conventional engineering, product and analytical responsibilities could blur. Middle-management layers could be removed.
The human employee would increasingly become an orchestrator, reviewer and decision-maker rather than the person directly performing every task.
This was not an isolated experiment.
Meta had already been building the technological foundations for agentic work. In April, the company introduced Muse Spark, describing it as the first model in a new series from Meta Superintelligence Labs.
By July, Meta was publicly describing Meta AI as a system that could plan, connect to applications, create presentations and perform tasks on behalf of users.
The internal workplace strategy and the external product strategy were therefore moving in the same direction: AI should not merely answer questions. It should act.
The 60% figure needs an important qualification
The most sensational part of Project OT was the possibility of cutting teams by 60%.
But that number requires precision.
Meta confirmed to Reuters that the exercise included scenarios in which some teams could be reduced by up to 60%. It was not a proposal to eliminate 60% of Meta’s entire workforce.
The distinction matters.
Meta had 77,986 employees at the end of March 2026, according to its first-quarter results.
A 60% reduction across the company would therefore have represented an extraordinary workforce elimination. That was not what Project OT ultimately proposed.
Instead, the internal exercise considered different levels of restructuring across different organizations. The mechanisms included layoffs, elimination of vacant positions, reductions in management layers and the redeployment of employees into AI-focused work.
Even so, the potential scale was enormous.
Meta had already gone through major workforce reductions during its earlier “Year of Efficiency.” Project OT represented a new philosophical justification for further reductions: AI would not simply make employees more productive; it could potentially change the number and type of employees required.
That is a fundamentally different proposition.
From AI assistance to AI replacement
For years, enterprise AI adoption was largely framed around augmentation.
AI could help a developer write code.
It could summarize documents.
It could analyze customer interactions.
It could draft marketing material.
It could search internal knowledge.
The employee remained responsible for the final work.
Agentic AI changes the equation.
An AI agent can potentially be given an objective rather than a narrowly defined task. It can navigate software, execute multiple steps, evaluate intermediate results and continue working without a human directing every action.
That is precisely why Meta’s internal strategy was so significant.
The objective was not merely to give employees better tools. It was to redesign the organization around the assumption that AI agents would become workers inside the workflow.
The company even began collecting workplace interaction data to help train such systems.
In April, Meta announced an internal Model Capability Initiative that captured mouse movements, clicks and keystrokes on designated work applications and websites, along with occasional screen snapshots. The purpose was to obtain real examples of how humans navigate computer interfaces so AI agents could learn to perform similar tasks.
The concept made technological sense.
The organizational implications were much harder.
Employees could reasonably ask a troubling question:
If the company is recording how I perform my job so an AI can learn to perform it, what happens to my job once the AI becomes good enough?
That question became central to the backlash.
When more code stopped meaning more productivity
The most important lesson from the Project OT episode may be hidden in the numbers.
At first glance, a 220% increase in code changes sounds like an extraordinary productivity breakthrough.
For a company betting heavily on AI coding agents, that could appear to validate the entire strategy.
But code is an intermediate output.
Customers do not buy lines of code.
They experience products, features, reliability, speed, security and usefulness.
According to internal information reported by Reuters, the increase in code changes to Meta’s internal platforms and infrastructure was accompanied by only a 36% increase in changes that resulted in new or upgraded features reaching users.
That difference is crucial.
Consider a simplified example.
If an engineering organization previously produced 100 units of code-related change and 20 units of meaningful product improvement, doubling or tripling code output would not automatically mean product value had doubled or tripled.
The additional output might create:
- More code to review
- More pull requests
- More testing requirements
- More integration work
- More technical debt
- More debugging
- More security exposure
- More infrastructure complexity
AI can therefore create an output inflation problem.
It can make the production of artifacts dramatically cheaper while making the validation of those artifacts more expensive.
That is exactly the problem enterprises need to understand as they move toward agentic development.
The hidden cost of autonomous agents
The second set of numbers was even more concerning.
Internal posts reportedly indicated that major technical and security incidents rose 40% compared with the previous year. Time spent dealing with those problems increased by as much as 70%.
The reported explanation was particularly revealing.
AI agents were capable of taking “large-scale, disruptive actions” that human employees would be unlikely to perform.
This illustrates a fundamental difference between human and agentic systems.
A human employee usually carries contextual knowledge about organizational norms, risk tolerance and consequences. They may stop because something “doesn’t look right.”
An AI agent can be extremely effective at pursuing the objective it has been given.
That is both its strength and its weakness.
If the objective is imperfectly specified, the agent may optimize the wrong thing.
If permissions are too broad, it may make changes across a much larger surface than intended.
If safeguards are inadequate, a small error can propagate quickly.
And if hundreds or thousands of agents operate simultaneously, the organization can experience failures at machine speed.
The result is a new category of enterprise risk:
AI can reduce the cost of execution while simultaneously increasing the cost of supervision.
Why Zuckerberg canceled the November wave
Against that backdrop, Zuckerberg’s decision to cancel the second restructuring wave becomes easier to understand.
The first wave was still executed on May 20.
Approximately 8,000 employees were laid off, representing about 10% of Meta’s workforce. At the same time, approximately 7,000 employees were moved into AI-related initiatives. Meta also eliminated around 6,000 planned openings.
But the second wave, planned for November, was canceled shortly before the first wave took place.
Reuters reported that the decision followed mounting employee resistance and internal evidence that AI-driven productivity gains were not developing as quickly as anticipated.
In other words, Meta did not abandon AI.
It abandoned the assumption that AI capability had advanced far enough to justify the next level of organizational reduction.
That distinction is critical.
The company continued investing heavily in AI.
What changed was the confidence that AI could immediately replace large amounts of human organizational capacity.
Meta Project OT AI workforce transformation became a lesson in the difference between output and value
This is where the Meta episode becomes much bigger than Meta.
Companies around the world are currently measuring AI adoption using metrics such as:
- Number of AI-generated code changes
- Number of AI interactions
- Hours saved
- Number of employees using copilots
- Number of automated workflows
- Number of tasks completed by agents
These metrics are useful.
But they can also be dangerously misleading.
A company can automate 80% of a process without improving the customer’s experience.
It can generate ten times as much content without increasing revenue.
It can produce three times as much software without shipping three times as many valuable features.
It can reduce headcount while increasing the workload of the people left behind.
The real measurement framework must therefore move further down the value chain.
AI productivity should ultimately be assessed through metrics such as:
Revenue generated → customer outcomes → product improvements → quality → reliability → risk → cost.
Activity belongs much further up the hierarchy.
The lesson from Meta’s internal numbers is simple:
More AI output is not necessarily more productivity.
The human cost was not limited to layoffs
The organizational transformation also created a significant morale problem.
Reuters reported that favorable employee sentiment in an internal Pulse survey fell from 74% to 55% during the period.
That is a dramatic decline.
The employee-monitoring initiative compounded the tension.
Meta subsequently scaled back elements of its tracking program after employee concerns. New controls allowed employees to pause collection for periods of time and request exemptions.
The controversy demonstrates another important issue with AI transformation.
Employees are not merely resources being optimized.
They are also the people responsible for implementing the transformation.
If employees believe an AI program is designed primarily to eliminate their jobs, they have an obvious reason to resist it.
This creates an organizational paradox.
The company needs employees to provide the data, expertise, judgment and feedback required to make AI systems better.
But an aggressive automation strategy can simultaneously convince those same employees that successful AI development threatens their careers.
That can undermine trust precisely when collaboration is most important.
The 7,000 employees moved into AI roles tell another story
It would be a mistake to interpret the May restructuring solely as a story of humans versus machines.
Approximately 7,000 employees were moved into AI-focused roles.
That is significant.
The workforce transformation was therefore partly about redistribution of human capital, not simply elimination.
Traditional roles could shrink while demand increased for people working on AI engineering, training data, evaluation, analytics, infrastructure and agentic systems.
This is consistent with a broader pattern emerging across technology companies.
AI may eliminate some tasks while simultaneously increasing demand for:
- AI infrastructure engineers
- Model evaluators
- AI safety specialists
- Data specialists
- Agent designers
- AI product managers
- Security engineers
- Human-in-the-loop supervisors
- AI governance professionals
The critical question is not simply how many jobs AI eliminates.
It is what happens to the composition of the workforce afterward.
Meta’s May restructuring provides an unusually clear example.
Thousands of jobs disappeared.
Thousands of employees simultaneously moved toward AI-related work.
That is transformation, not straightforward automation.
Meta’s AI spending makes the pressure understandable
There is also a financial dimension to the story.
Meta is spending extraordinary amounts of money building the infrastructure required for its AI ambitions.
In its first-quarter 2026 results, Meta projected full-year capital expenditure of $125 billion to $145 billion. By its second-quarter results, the company had narrowed the range to $130 billion to $145 billion.
That spending includes the infrastructure required to support Meta’s expanding AI capabilities.
The economic logic is therefore understandable.
AI infrastructure is enormously expensive.
If AI can simultaneously improve products and reduce operating costs, the return on that investment could be substantial.
But this creates pressure to demonstrate productivity gains quickly.
That pressure can encourage organizations to measure what is easiest to measure.
Code changes are easy to count.
Agent executions are easy to count.
Headcount reductions are easy to count.
Business value is harder.
Customer satisfaction takes time.
Product quality takes time.
Reliability requires sustained measurement.
Security failures can emerge months after deployment.
Organizational learning cannot always be captured in a dashboard.
That is why the distinction between activity and value is becoming one of the most important management questions in enterprise AI.
The irony: Meta is still becoming more agentic
Perhaps the biggest irony is that Project OT’s setback does not mean Meta has abandoned agentic AI.
Quite the opposite.
Meta’s public product roadmap shows continued movement toward AI systems that act rather than merely respond.
In July, Meta said Meta AI could create plans, connect to email and calendar applications, create slides and perform tasks on a user’s behalf.
That is exactly the technological direction that made Project OT possible in the first place.
The difference is that Meta appears to have discovered that technological capability and organizational readiness do not mature at the same speed.
An AI agent may be capable of completing a task.
That does not necessarily mean the organization is ready to give it unrestricted authority.
An AI system may generate code.
That does not mean an organization can eliminate the engineers who understand the architecture.
An agent may execute workflows.
That does not mean it can replace the people responsible for exceptions, accountability and judgment.
This distinction may become one of the defining principles of enterprise AI over the next decade.
What Project OT means for enterprise technology leaders
For CIOs, CTOs and business leaders, the Meta experience offers several practical lessons.
1. Do not equate automation with productivity
Automating a task reduces human effort.
It does not automatically increase business value.
Every AI initiative should therefore have a measurable connection to revenue, customer experience, quality, risk reduction or another strategic outcome.
2. Measure downstream outcomes
A developer writing more code is not necessarily more productive.
A customer-service agent handling more interactions is not necessarily delivering better service.
A marketing team generating more campaigns is not necessarily creating more revenue.
Measure the final outcome.
3. Treat AI agents as operational actors
Traditional software generally waits for instructions.
Agents can take actions.
That means agentic systems need identity, permissions, audit trails, rollback mechanisms, monitoring, testing and clear boundaries.
The governance model must evolve accordingly.
4. Keep humans where judgment matters
Human involvement should not automatically disappear simply because AI can technically perform a task.
High-risk decisions, security-sensitive operations, ambiguous customer cases and strategic decisions often require contextual judgment.
5. Do not optimize the organization before optimizing the workflow
Shrinking a team because AI is expected to replace its work can create a dangerous feedback loop.
The safer sequence is:
Understand the workflow → introduce AI → measure outcomes → identify bottlenecks → redesign roles → then change headcount.
Not the reverse.
6. Build trust into the transformation
Employees who understand how AI will change their roles are more likely to participate constructively.
Employees who believe AI is secretly being trained to replace them have a powerful reason to resist.
Transparency is therefore not merely an HR issue.
It is an AI implementation requirement.
Meta Project OT AI workforce transformation and the end of the “AI replaces everyone” narrative
The most interesting conclusion is not that AI failed.
It didn’t.
Meta continues to develop increasingly capable AI models, AI products and autonomous systems.
Nor did the restructuring disappear.
Meta still eliminated approximately 8,000 jobs and redirected thousands of employees toward AI-related work.
What failed was something narrower and more important:
the assumption that AI capability could advance quickly enough to justify a second, much deeper organizational restructuring.
The internal numbers exposed the problem.
AI could dramatically increase the amount of work being produced.
But increased work did not translate proportionally into increased value.
And in some cases, it increased the amount of work humans had to do to control, review and repair what AI systems produced.
That is the central lesson of Meta Project OT AI workforce transformation.
The future of enterprise AI is unlikely to be determined simply by asking:
“How much work can AI do?”
The better question is:
“How much valuable work can AI do reliably, safely and economically—and what human organization is required around it?”
Those are very different questions.
The bigger lesson for the AI economy
The Meta episode arrives at an important moment.
The technology industry has moved rapidly from AI copilots to AI agents. Companies are increasingly imagining digital workers that can execute multi-step tasks with limited supervision.
The temptation is obvious.
If an agent can do the work of five people, why retain five people?
But that calculation assumes that the work is fully captured by the task itself.
In reality, employees provide context, judgment, institutional memory, exception handling, quality control, risk awareness and accountability.
Those functions are often invisible in productivity calculations.
AI can automate the visible task while leaving the invisible responsibilities behind.
That is why the future workplace may not be a simple contest between humans and AI.
It may instead become a system in which humans define objectives, AI executes large portions of workflows, and humans remain responsible for outcomes.
The organizations that master that model could achieve enormous productivity gains.
Those that mistake AI-generated activity for genuine productivity may simply create faster ways to generate complexity.
Meta’s Project OT experiment offers one of the clearest warnings yet.
The lesson is not to slow down AI.
It is to become much more rigorous about where AI belongs, what it is allowed to do, how its output is measured and who remains accountable when it goes wrong.
In the AI-native enterprise, the scarce resource may ultimately not be computing power.
It may be human judgment.