AI spending is becoming a workplace behavior problem, not just a software invoice. Apple is placing a reported $200 daily cap on AI token use for engineers and may set a $25 allowance for non-engineers. Several agencies are already testing similar controls, although their approaches differ considerably.
The important shift is not the exact dollar amount. It is the decision to make AI usage visible, bounded, and discussable before costs become impossible to explain.
Caps are forcing better questions about AI use
Performance agency Roast has given employees set token allowances within its Claude subscription since March, with higher limits available during unusually busy periods. John Campbell, the agency’s head of innovation and AI, told Digiday that the system has discouraged people from using AI for repetitive tasks better handled by conventional software.
PMG takes a more explicit approach through its Alli4U portal, where staff face a $50 daily limit. Employees who repeatedly hit the ceiling can receive a larger budget or guidance on choosing a different model. That turns the cap into a human checkpoint rather than a simple blocked request.
Dept applies limits at team level, while Brainlabs uses tiers and asks employees to show their work when they reach a spending threshold. Good work can unlock the next tier. The underlying idea is consistent: AI access should remain flexible, but the cost of experimentation should not remain invisible.
The smarter limit may be judgment, not tokens
Hard caps create discipline, but they can also punish useful experimentation. Video and image production can consume more tokens than routine copy tasks, while media agencies testing agentic planning and buying tools may have very different needs from a social team drafting posts.
That is why some agencies are exploring “gating,” which routes a task from an expensive model to a cheaper one when the extra capability is unnecessary. Someone drafting a LinkedIn post probably does not need the most powerful model available. Training and model selection can therefore become a better guardrail than a universal daily ceiling.
The practical lesson for companies is to treat AI budgets like a design problem. A blunt limit may reduce waste quickly, but a tiered system can teach employees when AI is worth the cost and when another tool is enough. As model pricing and usage patterns evolve, the organizations that learn this distinction will have more room to experiment without allowing experimentation to become the strategy.