Companies that rushed to put generative AI tools in front of employees are now trying to slow their own workers down, according to reporting by 404 Media based on leaked Slack messages, internal dashboards, screenshots and emails from several companies.
The documents, obtained from half a dozen companies including Amazon, Adobe and Atlassian, show employers limiting access to AI systems, asking staff to use cheaper or less capable models, and in some cases cutting off particular models entirely. 404 Media also named Citi in its description of the affected companies.
The problem is not ideological. It is the bill.
404 Media reported that in at least one case, monthly AI spending had tripled to more than $15 million. The publication said the pattern spans tech, entertainment, banking and other industries, suggesting the issue is not confined to software companies experimenting with chatbots in a side channel.
Usage pricing is doing what usage pricing does
The mechanism is straightforward: enterprise AI tools often charge by consumption rather than by a predictable flat subscription. The more employees ask a model to read, write, summarize, code or analyze, the more tokens the company burns. Tokens are the chunks of text processed by these systems, and larger prompts, longer answers and more capable models tend to consume more of them.
That pricing model creates a different kind of software budget problem. A company can buy seats for a collaboration tool and know roughly what the bill will be. With metered AI access, a popular tool can become expensive fast, especially if employees treat the most powerful model as the default for routine work.
According to 404 Media, that is now prompting internal pressure campaigns and technical limits. Companies are not just encouraging staff to use AI less. They are steering workers toward weaker models, throttling usage and removing access to some systems when token spending gets out of hand.
Adobe is ending unlimited Claude access
One concrete example in 404 Media’s reporting is Adobe. The publication said emails show Adobe is ending unlimited access to Claude, Anthropic’s AI assistant. The reported change fits the broader pattern: companies may still want employees using AI, but they no longer appear willing to offer an open tap to expensive models.
The episode cuts against the tidy vendor pitch that AI adoption is an obvious productivity upgrade with costs that will sort themselves out later. The documents described by 404 Media show the later part arriving now, inside Slack threads, dashboards and internal emails where managers have to account for usage that keeps rising.
There is still plenty the reporting does not establish. It does not show whether the AI spending produced savings elsewhere, whether the restrictions are temporary, or how much employee productivity changed after access was reduced. It does show that some companies adopting AI at speed are now treating model access like any other metered infrastructure cost: something to monitor, cap and downgrade when the invoice gets ugly.
This story draws on original reporting from 404 Media.
