Companies Have Discovered the True Cost of AI, and They're Turning Off the Tap
The boom of generative artificial intelligence in businesses is beginning to show the bill to be paid. We wrote about it a few weeks ago, but the phenomenon seems to have an even more significant reach, highlighting how the introduction of AI in companies is by no means a simple operation and its returns are yet to be quantified.
According to a series of internal documents - Slack chats, emails, screenshots of dashboards - obtained by 404 Media from at least six major companies (including Atlassian, Adobe, Amazon, Citi, GitHub, and Accenture), various organizations are introducing restrictions on the use of high-end AI models to contain spending that, in some cases, has tripled in just a few months.
The most explicit case involves Citi Bank, which according to an internal email has disabled access to the most recent and expensive models - Claude Opus 4.6 and 4.7 from Anthropic and GPT-5.5 from OpenAI - since June 24, with reactivation planned for July 1. The internal communication specifies that these models "consume a significantly greater number of AI credits per interaction" and are the main cause of increased corporate expenditures. Employees have been asked to switch to less resource-intensive models: GPT-5.3-Codex for quick queries and simple code generation, Claude Sonnet 4.6 for code review and more complex architectural reasoning, reserving "premium" models only for truly necessary cases. The bank linked its decision to the switch of GitHub Copilot from a flat subscription model to a usage-based one, introduced in June. When contacted by 404 Media, Citi denied having disabled models or imposed token limits on employees, a denial contradicted by the documents cited by the same publication.
Atlassian, the company behind the Jira software, has eliminated unlimited use of AI tools by introducing a dashboard that shows employees the cost generated by their usage. According to data reviewed by 404 Media, monthly spending on services like AWS, Google Cloud, and OpenAI language models rose from 5 million dollars in August 2025 to over 15 million in May 2026, with an annual projection exceeding 120 million dollars. The company contested the accuracy of these figures without specifying which data is incorrect.
In Adobe, unlimited access to Claude will not be renewed and will expire on June 30, according to an employee. At GitHub, although there are currently no individual spending limits, the company is reportedly considering the adoption of open-source models to reduce costs and is testing user-based billing rather than team or project-based.
Amazon, meanwhile, has recently shut down an internal ranking that rewarded employees based on their level of use of AI tools, likely because it incentivized excessive and costly usage. Shortly after, some employees discovered the existence of previously unknown token limits. An Amazon spokesperson stated that guidelines on AI usage have not changed.
A similar picture emerges from other companies: an entertainment sector company has reached for the first time the monthly usage cap for ChatGPT, with a single developer responsible for nearly half of the total consumption, against a return on investment that is not clearly quantifiable. Accenture, for its part, has observed that much of the token consumption by clients does not stem from advanced software development activities but from simple operations like converting PDFs into presentations.