Atlassian, Citi, Adobe Cap Employee AI Access as Bills Climb

The most quietly interesting shift in enterprise AI this month is not a new model, it is a spreadsheet. 404 Media reported that Atlassian, Adobe, Amazon, Citi and GitHub are all pulling back on the unlimited-AI-for-every-employee posture that defined the last year of rollouts, because the bills got too big to ignore.

The concrete numbers give the picture. According to internal communications obtained by 404 Media, Atlassian's monthly AI spend went from around $5 million in August 2025 to more than $15 million by May 2026. Citi disabled access to Claude Opus 4.6, 4.7 and GPT-5.5 on June 24 and told staff in an internal note that "these models consume significantly more AI Credits per interaction" than mid-tier options, with re-enablement scheduled for July 1. Adobe is ending unlimited Claude access on June 30. Amazon quietly shut down an internal AI usage leaderboard and then, as one employee put it in a leaked message, went from "no more leaderboard to actual usage limits in two weeks." At an unnamed entertainment company, a single developer reportedly consumed almost half of the company's monthly ChatGPT token pool with, in a colleague's phrase, "no obvious ROI."

Why this matters for anyone doing this work at a real company: the pitch you have been sold, an unlimited seat to a frontier model, was a subsidy phase, and that phase is ending. The behaviour it produced is telling too. An Atlassian worker told 404 Media that colleagues "changed their workflow to maximize AI usage, and now they run out in 2-3 days," which is what happens when caps arrive after habits have already formed. Expect more employers to move to metered dashboards, to push work down to cheaper reasoning models, and to reopen conversations about self-hosted open-source options, which GitHub is reportedly already investigating.

The honest caveat is that this is one outlet's reporting drawn from leaked internal chatter at a handful of large companies, and it does not tell you whether the heavy top-tier usage was actually producing productivity gains that justified the spend. That is the real open question. The upside for practitioners is that the next quarter is going to force a much more careful conversation about which tasks genuinely need Opus 4.7 or GPT-5.5, and which have been reaching for them out of habit.