Meta weighed cutting teams 60% to go 'AI native' — then pulled back
Two views of the same bet landed today: an internal plan to hand Meta's workforce over to AI agents, and a fresh round of vendor promises that enterprise agents finally pencil out. Only one of them survived contact with reality.
Meta explored slashing many teams by as much as 60 percent in two waves to become an "AI native" company — then backed off after staff revolted and its own data showed the agents weren't delivering. That is the core of a Reuters investigation published Wednesday, built on internal documents and interviews. According to the report, internal data suggested the autonomous agent technology at the heart of the strategy was failing to produce the hoped-for productivity gains, while some investors had begun asking what Meta had to show for its enormous AI spending. The pullback tracks with Zuckerberg's own public admission in July that agent development is going slower than expected.
It is the sharpest evidence yet that "replace the org chart with agents" runs well ahead of what agents can actually do — and it comes from inside one of the biggest buyers in the market, not from an outside skeptic. Individual adoption keeps climbing; we noted last week that one in five US workers now hands tasks to AI, not colleagues. But a million small delegations are a long way from an AI-native org, and Meta apparently counted before anyone else had to.
Glean used its Glean:GO conference to unveil Tau, a desktop workspace that wires its enterprise-context AI into local files, apps and code — plus a set of benchmark claims aimed squarely at Anthropic. The San Francisco company said its testing shows a 5.2-times token-cost advantage per query over Claude Cowork, Anthropic's agentic work product, with users preferring Glean 3.6 times as often. We followed Cowork's rollout closely — Claude Cowork arrives in Chrome with skills and plugins. Tau, which plans multistep work and checks its own output on an open-source harness, is not generally available yet.
Treat the self-reported numbers with skepticism — vendors don't publish losing benchmarks. But the underlying pitch is where enterprise agent economics actually get decided: keep permission-aware company context resident, so the model stops paying to rebuild it with retrieval calls on every query. If that holds up independently, cost — not model quality — becomes the enterprise buying criterion.
What to watch: whether other large employers follow Meta's retreat from agent-first reorgs, and whether any independent lab reproduces Glean's token-cost numbers against Cowork.
If your employer announced an "AI native" restructuring tomorrow, would you believe the agents were ready? Tell us in the comments.
