Meta's Project OT imploded because the organizational design made the collapse inevitable.
The company committed at least $130 billion to AI chips and infrastructure in the same year it rebuilt its workforce around that bet [1] [2]. Layers of middle management were replaced by small pods reporting to unit heads responsible for 30 to 50 people, with performance ratings set by those unit heads, human resources, and unspecified AI systems [3] [2]. The internal Pulse survey fell from 74 percent favorable to 55 [4] [2]. Employees filled internal channels with protest and dark humor, clashed with the chief technology officer, and labor organizing gained traction [5] [2]. Code changes rose 220 percent while user-facing features rose 36 percent, incidents climbed 40 percent, and time lost to firefighting climbed 70 percent [6] [2].
The coverage split into a technology story about underperforming AI agents and a leadership story about a plan pushed too hard; this piece is the structural account of the operating system the redesign deleted, and of why the collapse was designed in.
The Redesign Deleted the Layer That Was Actually Running the Company
The coverage treated the deleted middle-management layer as overhead, and the redesign treated it as a line item; what the layer actually carried was the organization's implicit operating system – the judgment calls, the social accountability, the tacit coordination that made the formal structure work. In Harvard Business Review Prof. K. Sudhir names the mechanism: every organization runs on two operating systems, the formal one in the procedures manual and the implicit one that actually runs the place [7]. That implicit layer worked as background processes – invisible while running, noticed only in their absence. Delete the implicit system and the formal one keeps its shape while losing its function: the procedures still exist, and nothing makes them work.
The redesign uninstalled them and shipped no replacement.
The testimony from inside the pods describes an uninstall rather than a redesign. A Pod Lead reported receiving no manager training and no rating tools, while an internal AI-Native Playbook had circulated since October 2025 and eleven units had moved onto pods by June [8] [2]. Unit heads took on responsibility for 30 to 50 people – a span at which social accountability stops functioning – with ratings set by unit heads, human resources, and unspecified AI systems, while roughly 8,000 jobs were eliminated in the first wave and a second wave was planned for November [8] [3] [2]. The functions the deleted layer had carried – translating ambiguity between strategy and execution, holding judgment, keeping coordination running quietly underneath the formal structure [9] [10] – had nowhere to go.
The organization registered the loss the way any system registers missing background processes. Informal hierarchies began quietly rebuilding around the flat structure – snack quality, offsites, fewer reports per manager became the new status markers [11]. These are gestures, and they are also something more telling: status hierarchies are coordination infrastructure, and when the formal version is deleted, an informal version grows in its place [12]. The system was already compensating for what the redesign removed.
The New Structure Concentrated Authority at the Center and Distributed Responsibility at the Edge
The deleted functions had to land somewhere, and the new structure decided where: decision rights locked at the center, information and responsibility pushed to the edge. In operating-system terms, the redesign rewired permissions – and the rewiring ran in one direction. Pods displayed the interface of distributed authority while the underlying permissions stayed centralized. Form without substance.
The official-versus-internal record is the exhibit. Meta's public position holds that rating and promotion decisions "were and are made by people" [13]. The internal record describes ratings set by unit heads, human resources, and unspecified AI systems [3] [2], day-to-day priorities determined by "agent-assisted analysis" [14], and an applied AI engineering group running at about 50 engineers per manager [8]. Both descriptions can be accurate at once – that is the point: a structure can distribute the appearance of authority while keeping every consequential permission at the center.
Pods displayed the interface of distributed authority while the underlying permissions stayed centralized.
The arithmetic of the redesign shows the same inversion. Unit heads carrying 30 to 50 people, managers carrying 50 – these are ratios, and ratios alone do not drive performance [15] [16] [17]. What drives performance is redesign: which decisions move where, which information travels with them, who can act without asking. Organizations that distribute authority intentionally – Handelsbanken's branch autonomy, HolacracyOne's explicitly defined decision domains – design the permissions first and the structure second [18] [19]. Meta's redesign worked in the opposite order: it changed the structure and assumed the permissions would follow.
The structure was now complete, and it had a property the coverage missed: it concentrated authority over tension at the center while distributing the experience of tension to the edge.
The Structure Generated Tension and Built Nothing to Process It
A structure that concentrates authority and distributes responsibility guarantees a specific output: tension at the edge with no channel to resolve it. A designed organization metabolizes that tension through dialogue, through decision-rights redesign, through structures that let disagreement become a decision rather than a grievance. Meta's structure shipped suppression instead, and the replacement included no error handling; warnings queued as unhandled exceptions until the queue discharged.
The suppression mechanisms are documented. Employees worked under replacement threats while a tool recorded every keystroke and mouse movement [5] [2], deployed without an opt-out [5]. The morale gestures followed: snack quality, offsite budgets, fewer reports per manager [5] [11]. The gestures managed the appearance of tension while the tension itself accumulated.
The discharge arrived in the numbers and the language. The Pulse survey fell from 74 percent favorable to 55 [4] [2]. Employees filled internal channels with elephant imagery [5] [2] and coverage of the tracking tool coined the phrase "Employee Data Extraction Factory" [5]. The chief technology officer described morale as almost the worst it had been in 20 years, and the company's response included snacks [5].
The responses stayed gestural while the pressure kept building. Labor organizing gained traction [2]. The keystroke tracking paused only after the revolt [8]. In July, Zuckerberg conceded that agent technology "hasn't really accelerated in the way that we expected" [11].
We read the revolt as accumulated tension finding its only available outlet. The pattern extends beyond Meta: 29 percent of employees admit to actively sabotaging AI initiatives, and 54 percent of C-suite executives describe AI adoption as corrosive to their organizations [20]. Research links AI adoption to employee harm through degraded psychological safety [21]. Across a 39,000-worker survey spanning 36 countries, the workers closest to the work carried the fear while leadership carried the optimism [22]. The tension was documented, dated, visible in every channel the structure left open – queued from the start.
The replacement included no error handling; warnings queued as unhandled exceptions until the queue discharged.
The Deleted Functions Returned as Behavior
The queued tension had to go somewhere, and the structure had already decided where: into behavior. The revolt, the sabotage, the decline – these are products of the design, not causes of the failure. The deleted coordination functions returned as behavior, amplified by AI agents executing whatever ambiguity the structure encoded.
The telemetry shows the return. AI agents took "large-scale, disruptive actions" humans were unlikely to make [23] [6]. Code changes rose 220 percent while user-facing features rose 36 percent; incidents rose 40 percent and time lost to firefighting rose 70 percent [6] [2].
The agents executed the ambiguity well: with no one carrying judgment and no structure processing tension, they optimized locally – and the collective result was systemic drift. Reassigned engineers called themselves "draftees", set to writing puzzles that would train the models meant to replace them, and coverage of the reassignments reported engineers calling the AI unit a "Gulag" [11]. Every one of these outputs traces to the structure, and the internal texture is reported by the Reuters investigation and its coverage [2].
The most consequential output was: before the implosion, the structure produced silence – and leadership misread it as alignment [24] [25]. Research on organizational silence documents the mismatch: leaders interpret withheld voice as consent, while employees withhold because the structure gives disagreement nowhere productive to go [25]. The silence was the system running without its operating layer; the revolt, the sabotage, and the telemetry together read as the crash log of the same system, later in the same run.
The deleted coordination functions returned as behavior, amplified by AI agents executing whatever ambiguity the structure encoded.
The outputs were legible to anyone reading at the system level. The question the case raises is why the readings stopped at the component level.
The Market Diagnosed the Outputs and Left the Structure Untouched
The component-level readings generated component-level responses. July brought concessions – snack quality, offsite budgets, fewer reports per manager, and the "betting on people" campaign [11] [13] [2]. The dual-cause reading (the revolt plus the failing agents) [26] refined the component story without leaving the component level. The crash report was read at component level – bad technology, bad operator – and answered with component swaps. July was a patch release, and patches leave the underlying installation untouched.
The $130 billion authorization never had to answer a structural question [1]. The governance question would be: Who in the room holds the authority to ask whether the organization's operating structure can carry the bet? The documented pattern makes the question urgent.
Only eight percent of boards hold AI oversight at the board level [27]. Analysts describe boards facing an accountability reckoning on AI decisions [28]. Governance commentators now argue that AI governance is a fiduciary duty [29]. The question is open at Meta; the pattern says it is open almost everywhere.
The industry pattern around the case confirms the reading. 95 percent of organizations report zero measurable return on AI investment [30]. Ford made the same category of deletion and corrected it by rehiring 350 veteran engineers [31]; 55 percent of employers admit their redundancy decisions were wrong [32], and the reversal pattern now runs across industries [33] [34] [35].
Reported observation elsewhere: Oracle ran the same dynamic – managers identifying employees for cuts while headcount fell by roughly 21,000 year over year [36]. The distinction between the two course corrections is instructive: Ford reinstalled the deleted layer. Meta patched the outputs and left the structure as designed.
July was a patch release, and patches leave the underlying installation untouched.
Every redesign is an uninstall. The organization that replaces the layer with nothing discovers, by collapse, what the layer was carrying – the judgment, the accountability, the tension processing that made the formal structure operable. Meta's implosion was the system performing as designed: authority concentrated, tension suppressed, functions returning as behavior, outputs patched. The choice for any organization approaching the same redesign is a replacement operating system in hand, or discovering the need for one by collapse.
Sources
- Katie Paul & Jaspreet Singh (Reuters; via GBAFR) — Meta lifts capital expenditure forecast, doubling down on AI push
- Katie Paul (with Jeff Horwitz; Reuters Investigations) — Mark Zuckerberg had a bold plan to replace Meta staff with AI. Here's how it imploded
- Editorial Desk (ET Enterprise AI) — Inside Meta's 'Project OT': 60% job cut scenario, internal revolt and a sudden U-turn
- Storyboard18 (editorial desk) — Meta explored cutting some teams by 60% for its AI push, then Mark Zuckerberg pulled back: What is Project OT?
- Anna Iovine (Mashable) — Meta faces employee backlash over tracking tool
- Evan Schuman (Computerworld) — Meta's plans to replace workers with AI fell flat, report says
- K. Sudhir (Harvard Business Review) — How to Design Agentic Systems Around the Implicit Rules that Govern Your Company
- Camila Nogueira (BetaNews) — Zuckerberg's secret AI plan to remake Meta's workforce collapsed
- Kara Dennison (Forbes) — Why Companies Cutting Middle Managers To Fund AI Is A Mistake
- McChrystal Group — Middle Managers as Strategic Accelerators: Reframing the "Frozen Middle"
- Vytautas Valinskas (Technology.org) — Meta scrapped its plan to replace staff with AI
- Betterworks — The Great Flattening: What Happens When Middle Management Disappears?
- Nucleus_AI (YourStory) — Meta tried to replace workers with AI. It didn't go as planned
- Will Shanklin (Engadget) — Meta reportedly abandoned an AI-focused restructuring plan that would have laid off thousands
- Pawan Joshi — Span of Control Math
- Catherine Scott (AIHR) — HR's Guide to Calculating Span of Control
- Bauerlein, J., Fitzpatrick, T.A., & O’Neill, S. (Kaufman Hall) — The span-of-control myth
- Don Ledingham (Substack) — Handelsbanken – a case study in the power of decentralisation
- HolacracyOne — Distributed Authority
- Writer Team (Writer) — Enterprise AI adoption in 2026: Why 79% face challenges despite high investment
- Byung-Jik Kim, Min-Jik Kim, Julak Lee (Nature HSSC) — The dark side of artificial intelligence adoption: linking artificial intelligence adoption to employee depression via psychological safety and ethical leadership
- Nick Lichtenberg (Fortune) — Workers around the world are scared. A massive new survey shows just how much
- Scharon Harding (Ars Technica) — AI agents meant to replace Meta workers made 'large-scale, disruptive actions'
- Benjamin Laker (Forbes) — Why Employees Stop Speaking Up And Leaders Miss The Warning Signs
- Andy Cleff — Lifting the Curse of Organizational Silence
- Maximilian Schreiner (The Decoder) — Employee revolt and failing agents forced Meta to scrap its AI layoff plan
- Alston & Bird — How Boards Can Shrink the AI Governance Gap
- ComplexDiscovery Staff — Governing the Ungovernable: Corporate Boards Face AI Accountability Reckoning
- Patrick Meson / Kevin M. LaCroix (D&O Diary) — AI Governance Is a Fiduciary Duty
- Legal.io (editorial) — MIT Report Finds 95% of AI Pilots Fail to Deliver ROI, Exposing "GenAI Divide"
- Anthony Ha (TechCrunch) — Ford rehires 'gray beard' engineers after AI falls short
- Orgvue — 55% of businesses admit wrong decisions in making employees redundant when bringing AI into the workforce
- Justina Lee (CNBC) — Employers who laid off workers citing AI are already starting to regret it
- Joe Toscano (Forbes) — Ford Hiring 350 Engineers After AI Failed Shows Human Value In AI Era
- Matthew Sellers (HCMAG) — Ford's 'gray beard' rehire is a warning shot for AI-first workforces
- The Grind Hotline (Substack) — Meta's secret AI layoff plan exposed
Please note: 51even is an AI-first organization. We embrace AI at every step of our value creation and build our processes with a deep integration of human-AI capability. Humans always have the last decision. But this text was heavily built with AI.