Meta intended to reduce certain teams by 60% and substitute their roles with AI.

      Meta dedicated the first half of this year to planning the replacement of much of its workforce with AI agents, but ultimately abandoned the most ambitious aspects of the initiative when the technology failed to meet expectations. This initiative, referred to as Project OT (short for Organization Transformation), aimed to reduce the size of many teams by as much as 60%.

      Katie Paul provided details of the plan to Reuters, utilizing a wealth of internal documents, online posts, recordings, and interviews with over 20 individuals familiar with the company.

      Mark Zuckerberg and his top executives devised the plan at his Hawaii residence in January, envisioning AI taking over routine tasks performed by thousands of employees. Smaller groups, described in one planning document as “talent-dense” teams, were expected to oversee these virtual workers.

      The plan was executed in two phases. The first wave of layoffs was set for May, with a second round planned for November. Both layoffs and the closure of open positions were intended, alongside the dismissal of employees deemed underperformers. An internal document indicated that the layoffs could be as large as or exceed the approximate 25% reduction Meta enacted three years ago.

      However, on the night of May 19, just hours before the first wave of layoffs, Zuckerberg decided to cancel the November phase of the plan. The following day, Meta proceeded with cutting around 8,000 jobs, nearly 10% of its workforce, while appointing 7,000 employees to AI-related roles that same week. Reuters could not determine why Zuckerberg changed his mind.

      Meta confirmed the existence of Project OT, characterizing it as a year-long initiative aimed at reducing costs, reconfiguring team structures, and reallocating employees to priority areas such as training data production. The company also acknowledged the two-stage structure and the 60% reduction figure, clarifying that it applied only to particular teams and that many major units were excluded from the plan. The second wave was canceled before any specific job reductions were finalized.

      “This ultimately resulted in transitioning thousands of employees to engage in priority tasks on several newly-formed teams, as publicly reported. We did not proceed with every scenario from the exercise, nor was it ever assumed we would,” a Meta spokesperson stated.

      The idea originated from Meta executives studying how AI startups structure their organizations last year. Chief data officer Alex Schultz and head of product Naomi Gleit both traveled to Asia, where sources indicated they were impressed by how these startups organized their teams around AI.

      Gleit informed Reuters in June that she had spent considerable time in Meta’s Singapore office and that practices there had “inspired some of the teams in California and New York.” Many ideas were reportedly generated from the ground up.

      However, Meta’s internal data indicated that the AI agents were generating quantity over quality. According to a June post by chief technology officer Andrew Bosworth, code modifications to internal systems surged by 220% year-over-year, while features reaching users only rose by 36%.

      This discrepancy became problematic. An April post noted that uncontrolled AI agents were undertaking “large-scale, disruptive actions that humans likely would not perform.” Significant technical and security incidents, including service outages and potential data breaches, increased by 40% compared to the previous year, with time spent addressing these issues rising by 70%.

      Infrastructure teams had raised reliability concerns as early as March. In June, hackers exploited Meta’s new AI customer support bot to access high-profile Instagram accounts, including the dormant Obama White House page.

      Meta declined to comment on the internal data.

      Employees began to believe they were training their replacements. Meta had assigned engineers to create software challenges used as training data, leading many to describe the assignments as rote in internal communications. The company also implemented tracking software on devices belonging to US employees to monitor keystrokes and mouse clicks, allowing agents to learn computer usage behaviors.

      Meta put a halt to this program in June. Since then, 26 employees have filed a lawsuit against the company, alleging that its AI systems targeted employees on medical leave for layoffs.

      In response to executive communications, staff shared images of elephants, reflecting their sentiments. Meta's internal sentiment score dropped from 74% favorable to 55% according to its half-year Pulse survey, leading to increased labor organizing efforts.

      The proposed new structure involved a shift away from traditional organizational charts. An internal document titled the AI-Native Playbook outlined the new structure, proposing that product teams of 10 to 20 specialists would be reduced to smaller groups of three to five. Engineers, designers, and product managers would hold a single title of “builder,” with the elimination of middle management layers, while daily priorities would be determined by “agent-assisted analysis.”

      By June, at least 11 units had begun to adopt this pod structure. Unit leaders would manage between 30 to 50 employees, supported by human resources personnel and unspecified “AI systems.” Meta did not clarify this term, but emphasized that performance evaluations and promotions were still conducted by humans, not AI.

      Z