After months of encouraging AI experimentation, companies like Atlassian, Adobe, Amazon, and Citi are restricting employee access to expensive models as internal token costs spiral upward.

Major companies are pulling back on employee AI access after experiments with unlimited model use produced runaway costs, according to internal communications and employee accounts obtained by 404 Media and corroborated by reporting from heise online.

Atlassian, which had encouraged maximum AI utilization, recently stopped unlimited use and launched an internal cost-tracking dashboard. Internal data shows AI spending surged from $5 million monthly in August 2025 to over $15 million by May 2026, with annual projections exceeding $120 million. The company denied the figures when questioned by 404 Media, but did not specify which internal numbers were incorrect.

Adobe told employees that unlimited access to Claude would expire on June 30, 2026, with workers advised to complete AI-dependent tasks before the cutoff. Amazon introduced token limits just weeks after deactivating an internal AI usage leaderboard that had encouraged maximum utilization. "Crazy, we go from no more leaderboard to actual usage limits in two weeks," read an internal Slack exchange quoted by 404 Media.

Citi temporarily blocked employee access to newer models from Anthropic and OpenAI, directing staff to use specific models for particular tasks. The bank denied the action despite internal emails proving the block, heise online reported.

The reversals come as companies face per-token pricing models that can burn through budgets quickly. One unnamed media company hit its monthly ChatGPT limit with nearly half the usage attributed to a single employee who produced no obvious return on investment, according to 404 Media.

The cost-control measures contrast with public encouragement of AI experimentation. None of the companies disclosed specific AI spending figures in their public filings, leaving investors and employees to piece together the economics from internal documents and leaked communications.