The US Army told its staff in May 2026 that they had unlimited AI access. By mid-June, the pool was empty. The Army's AI platform, Ask Sage, runs on a token-based subscription: the organisation buys a block of usage upfront, and every question asked, every document drafted, every task sent to the AI draws from that block. The Army had 100 million tokens for the year. Staff burned through them in weeks.
To put that in perspective: a single back-and-forth conversation with an AI tool can consume a few hundred tokens. A power user doing serious document work might use 200,000 tokens in a month. Scale that across a force of 3.5 million people who were actively encouraged to use the tool, and the budget collapses fast.
The Army is not the first organisation to learn this lesson the hard way. Uber rolled out an AI coding tool to its engineers in late 2025. By April 2026, the company's entire annual AI budget was gone, four months into the year. Uber has since capped each engineer at $1,500 per month per tool. Meta built an internal leaderboard ranking all 85,000 employees by how many AI tokens they consumed, treating high usage as a sign of productivity. Costs headed toward billions. Meta shut the leaderboard down and started steering staff toward its own cheaper in-house tool. Microsoft cancelled a large block of AI coding licences for similar reasons.
What made all of this worse is that most organisations had no real-time view of the spending as it happened. A KPMG survey of more than 2,000 senior business leaders across 20 countries found that only 26% of large companies have full visibility into their AI costs. Another 22% only found out how much they had spent when the bill arrived. Nearly half said they had already pulled back or stalled AI deployments after realising costs had exceeded the value they were getting.
The Army situation has an added layer. The Department of Defense burned through 20 billion tokens per day during its 38-day air campaign against Iran, using a separate military planning system called Maven. That is a completely different scale of use: operational, real-time, tied to active military targeting. The question of whether that usage draws from the same budget pools as regular staff using Ask Sage for paperwork is still unanswered publicly.
There is also a performance question. One Army employee told Wired they did not find the tools particularly useful for their work, and that one model claimed to have completed a task it had not actually done. That is a real problem in any setting, but especially in one where the outputs might feed into decisions that matter.
The broader lesson for any business operator watching this is straightforward. AI is not priced like software. A traditional software licence costs the same whether your team uses it heavily or barely at all. AI tools that charge per token cost more the more your staff use them, and usage tends to grow faster than anyone expects once access is easy and free. Encouraging heavy use without a spending cap is the same as opening a company card at a hotel minibar and telling staff to help themselves.
Organisations that are getting real value from AI share one trait: they track what it costs at the task level, not just the total bill. They know which workflows justify the spend and which ones are just staff experimenting. The KPMG data confirms it: companies with strong cost visibility are five times more likely to report a measurable return on their AI investment.
The Army will almost certainly renew its subscription. The Pentagon has a $29.5 billion AI infrastructure budget request for the next fiscal year. But running out of tokens six weeks after promising unlimited access is a signal that even the organisations most committed to AI adoption have not figured out how to manage it as a financial operation. That gap is where the real work is.