Meta Pulls Back on AI Plan as More Code Fails to Produce More Features

SAN FRANCISCO — Meta Platforms Inc. has recalibrated an internal initiative known as “Project OT” after the company’s effort to replace corporate management and software engineering workflows with autonomous artificial intelligence systems encountered operational limits. Details of the project were reported by Reuters.

Project OT was conceived as an “AI-native” corporate architecture built around autonomous software agents. The blueprint called for agents to handle work previously performed by thousands of human employees, flattening organizational hierarchies and shrinking engineering and product teams. Meta Chief Executive Officer Mark Zuckerberg also began using a customized “CEO agent” designed to aggregate internal enterprise data and answer high-level strategic and operational questions directly.

The system allowed Zuckerberg to query company systems without relying on the traditional multi-tier reporting structures that gather, interpret, and escalate internal information. In traditional corporate governance, those layers can challenge the premises of executive questions, identify historical precedents, and flag edge cases in which data may be misleading.

Meta’s retreat came after internal evaluations raised doubts about the real-world returns of Project OT’s most aggressive workforce-reduction targets. AI-assisted software tools sharply increased the volume of computer code written by engineers, but the increase in raw output did not produce comparable gains in user-facing features or product updates across Facebook, Instagram, and WhatsApp.

The initiative followed a broader restructuring that began after rapid pandemic-era hiring. In late 2022, Meta laid off 11,000 employees, or approximately 13 percent of its workforce. In March 2023, Zuckerberg declared a “Year of Efficiency,” followed by another 10,000 job cuts and the cancellation of roughly 5,000 open roles. Those measures were aimed at stripping away middle management and reducing organizational overhead.

Meta simultaneously invested tens of billions of dollars in specialized hardware infrastructure. The company acquired hundreds of thousands of Nvidia H100 Tensor Core GPUs to build massive compute clusters totaling the equivalent of 600,000 H100 GPUs, while developing and open-sourcing its Llama family of large language models as it positioned itself as a central platform provider in the generative AI race.

Project OT was intended to use those internal models for routine corporate mechanics, which organizational researchers define as “coordination friction.” The category includes meeting scheduling, project tracking, cross-departmental data reconciliation, documentation, and routine approvals. Automating those logistical steps can eliminate delays and streamline information flow.

Software engineers and management experts distinguish that administrative work from “cognitive friction”: deliberate pushback, competing interpretations, and historical context introduced when diverse teams examine difficult problems. AI tools may eliminate administrative drag, but automated systems can struggle with the scrutiny that comes from disagreement and independent evaluation.

A corporate AI agent trained on internal data can also become a confirmation engine. Because such models adapt to an individual user’s historical prompts, preferences, and established assumptions, an executive-level agent may synthesize information in ways that reinforce the executive’s existing worldview. Instead of serving as an independent evaluator, it can package internal data into articulate, self-referential responses while filtering out critical dissenting perspectives before they reach leadership.

Corporate governance analysts refer to this dynamic as automated self-referentiality. An executive may experience the illusion of fast, independent intelligence while interacting with a system optimized to match that executive’s operational preferences.

Meta’s results resemble broader findings from the software industry. GitHub Copilot and internal generative coding models can accelerate initial code generation and syntax completion, yet engineering organizations increasingly report bottlenecks in code review, system integration, architectural design, and quality testing. The difference between code production and product delivery at Meta leaves raw software output disconnected from functional innovation and enhanced user experience.

Comparable concerns are appearing in scientific and industrial research. Fully integrated end-to-end AI systems are being deployed to formulate hypotheses, design experiments, analyze data, and draft technical papers. These automated loops can dramatically accelerate routine research, while domain experts warn that systems trained on existing datasets tend to optimize within established parameters and may reduce the experimental variation needed for breakthrough discoveries.

The technology sector is debating the distinction between administrative speed and strategic decision-making as companies integrate generative AI into core corporate operations. For Meta and the broader tech industry, the lessons drawn from Project OT mark a transition point in enterprise AI deployment: as tech executives continue funding heavy infrastructure investments, the challenge is ensuring that AI systems preserve critical oversight, dissent, and independent evaluation within corporate governance.