A single number captured the mood of enterprise AI this week: $500 million, the amount one company spent in thirty days on Anthropic's Claude after giving employees unrestricted access. The figure, reported by Axios and relayed by Business Today, is extreme, but it works less as an outlier than as a symbol. Across the industry the bills are arriving faster than the returns, and a year of breathless adoption is curdling into something more sober.
The spending problem now has a name: "tokenmaxxing," the habit of staff generating ever more prompts and automated workflows until costs spiral past any plan. Google's Sundar Pichai put it plainly, noting that companies "are already blowing through their annual token budgets and it's only May." The behavioural traps are everywhere. Amazon, according to a Financial Times report, shut down an internal leaderboard meant to encourage AI use after employees began prompting purely to climb the rankings, running up compute bills for no work in return. Microsoft is winding down internal Claude Code licences by the end of June, and Uber's operating chief says the costs have become "harder to justify." The instinct to just keep prompting has met a budget that cannot keep pace.
The harder question is whether all that spending buys anything. Gartner now places generative AI in what it calls the "Trough of Disillusionment," the stage where inflated expectations collide with real-world outcomes, while AI agents still sit at the giddy peak just before it. The Register reports Gartner's blunt forecast that most generative and custom-model projects will end as busts, and a widely cited MIT study found that 95 percent of corporate generative-AI pilots are failing to deliver. The recurring diagnosis is not that the models are weak but that the organisations around them are not built to use them: data sits in silos, ownership is unclear, and governance is patchy or absent.
The most striking new evidence, though, is behavioural. A Boston Consulting Group study of more than 1,200 finance and HR professionals found that when a flawed document was attributed to a named AI "employee" rather than to a person, reviewers caught fewer of its errors, took less personal responsibility for them, and were quicker to push the work onto a colleague to double-check. Nearly a third of managers now describe AI as a teammate, and more than a fifth list AI agents on their org charts. The lead researcher, Matthew Kropp, calls the effect "passing the buck": treating the tool as a coworker lets humans offload accountability onto software that cannot hold any. Far from easing adoption, the practice left workers 10 percent less trusting of how AI would be deployed, and more worried about their own jobs.
Put together, the week describes a double reckoning. One is financial: the meter runs whether or not the work improves. The other is organisational: handing powerful tools to teams without measurement, guardrails or clear lines of responsibility produces churn rather than productivity. The market is still vast. Menlo Ventures pegs enterprise generative-AI spending at $37 billion in 2025, up from under $2 billion two years earlier. But the easy phase is ending. The companies that come out ahead will be the ones that can answer a question most still cannot: not how much AI they are using, but what, precisely, it is buying them.