Corporate investment in AI is shifting from encouraging usage to controlling costs. Several major tech and retail companies have recently restricted employees' access to or usage limits of AI tools, as agent-based products repeatedly invoke models, causing corporate bills to rise rapidly.
According to reports from The Information, Bloomberg, and the Financial Times, AT&T has restricted some employees’ access to GitHub Copilot, owned by Microsoft; Meta has also tightened spending on Anthropic and other AI services. Uber, Walmart, and Amazon have similarly adjusted internal policies, all focusing on reducing AI usage costs.
Companies are beginning to set limits on AI
Cost pressures have spread from individual teams to the company level. The report notes that for companies with high AI usage, related expenses per employee per month can reach $7,500. Even as the cost per model invocation decreases, agent tools frequently invoke the model multiple times during task execution, ultimately driving up the total bill.
Uber is a prominent example. The company exhausted its full-year AI programming budget by April 2026 and subsequently set monthly usage limits of $1,500 per employee and per tool. Walmart has also implemented usage caps for its internal AI agents. Amazon eliminated its internal leaderboard that ranked employees based on AI usage, in order to reduce unnecessary consumption driven by ranking competition.
Microsoft also found that some engineers spent between $500 and $2,000 per month solely on Claude Code. This highlights that enterprise AI costs are not just about purchasing models, but are directly tied to how they are used.
The disagreement centers on efficiency versus budget.
Not all companies have chosen to simultaneously tighten controls. Aaron Levie, CEO of Box, said the company never used leaderboards to encourage employees to consume tokens excessively, so no such bias emerged. The head of engineering at Databricks stated that the company currently does not set AI budget caps for engineers.
This reflects differing internal judgments within companies. Some organizations believe that restricting usage helps bring costs back within budget as quickly as possible, while others believe that as long as employees can convert AI into higher productivity, continued investment remains worthwhile.
However, the core issue remains unchanged. Previously, companies were willing to increase AI spending based on expectations of productivity gains; now, significantly reducing usage could also diminish these returns.
Demand for cost management tools is increasing
Tighter budgets are also shifting the beneficiaries within the AI industry chain. An increasing number of companies are moving simple tasks from expensive, cutting-edge models to more affordable or open-source alternatives, reducing expenses without significantly lowering usage.
Under this backdrop, demand is rising for model gateways, usage monitoring, and model routing tools. Microsoft and Databricks have launched related gateway products to help businesses monitor employee usage and set spending limits. This month, Factory, an AI software company backed by NVIDIA, also released a model router designed to automatically assign low-complexity tasks to less expensive models.
Microsoft CEO Nadella recently stated publicly that AI models should operate like interchangeable commodities, rather than concentrating value in the hands of a few model providers. This stance aligns closely with current enterprise customers' demands for cost reduction.
Microsoft emphasizes controllable costs when launching new products.
In response to customers tightening their budgets, Microsoft this week revealed the pricing for its new product, Copilot Cowork. The product, primarily powered by Anthropic models, is designed to automate more complex, multi-step tasks within Office 365.
Its pricing model is “license fee plus pay-as-you-go.” Users must first purchase a 365 Copilot license, starting at $30 per month, and then pay additional fees based on actual usage of Copilot Cowork. This closely resembles Anthropic’s earlier billing approach for enterprise customers.
Microsoft also emphasized that customers can set usage limits for Copilot Cowork and plans to support switching between OpenAI, Microsoft’s proprietary, or other models within the product to reduce costs. The report also stated that Microsoft is testing solutions to replace Anthropic models with open-source models in certain scenarios.
This indicates that the new competitive frontier in the enterprise software market is shifting from merely comparing model capabilities to who can maintain performance while keeping AI costs within an acceptable range for businesses.