Meta may cap employee AI token usage as AI-driven costs rise

Meta spent much of 2026 pushing toward a workplace built around AI agents and smaller human teams. Now, the company is confronting a less obvious problem in its strategy. The more its employees use AI, the more it can drive up Meta's expenses.

In response, Instagram chief Adam Mosseri said in July that Meta could eventually need to set limits on the amount of AI tokens individual engineers can consume. He was discussing a possible change rather than announcing a new policy rollout. Meta does not currently have employee token caps, but Mosseri said they could be introduced as AI usage grows.

However, following Project OT and Meta's cancellation of the second round of layoffs, a possible future policy to curb AI usage underscores how complicated creating an ideal AI workplace can be.

Meta's AI workforce experiment went further than simple automation

Meta's AI ambitions go well beyond using it as a productivity tool. First revealed by a Reuters article, the Project OT initiative, short for Organization Transformation, explored reorganization of teams around AI agents, with some teams potentially reduced by up to 60%.

The first round of layoffs occurred in May, when Meta cut about 10% of its workforce. But Meta pulled back right before its planned second round of layoffs was expected to happen. This reversal highlighted the potential problem with treating AI as a direct replacement for human labor.

Making more AI available or spending more tokens does not automatically mean the work can be accomplished with fewer people. The reporting on Project OT found problems with AI reliability, employee resistance, and the practical difficulty of reorganizing teams around AI systems that still required quite a bit of human oversight.

Meta has since acknowledged the existence of Project OT but stated that it was a scenario planning exercise and that the company did not move forward with every scenario.

Now Meta is confronted with the cost of using too much AI

Every request to an AI model requires computing resources. These costs are often measured using tokens, also known as units of text that AI systems process when they read instructions and generate responses. The more complicated or frequent the request is, the more tokens it can require.

Speaking about Meta's token costs on the Lenny's Podcast, Mosseri said that currently there are no limits on token spending. However, he could imagine a point within a year or two when Meta may need to establish some spending limits, especially if AI usage goes up.