The Wall Street Journal published a number on Sunday evening that I am going to remember: out of nearly 400 businesses surveyed, 11% could accurately forecast their AI spending. Not hit the figure exactly. Just forecast it.
First #tokenmaxxing, then the bill
The WSJ sums up the backstory in two sentences. American companies first told their staff to use as much AI as they possibly could, hashtag #tokenmaxxing. Then the bill arrived, and those same companies started counting tokens.
The reason sits inside the billing unit itself. How many tokens a model burns on a given task is not knowable in advance. Two employees ask the same question; one gets a short answer, the other watches the model think for three minutes and call four tools. Both are correct behaviour. The invoices differ by a factor of ten.
Why tokens resist budgeting
Anyone working with Claude Code or an agent knows this from their own sessions. Cost does not track the number of tasks. It tracks how long the road to the answer was. An agent that wanders off and restarts twice costs a multiple of the same agent that turns the right way on the first try. The longer sessions get, and the more steps a run takes on its own, the wider that spread.
For a finance team that is an awkward property. Server cost scales with load, licences scale with headcount, and both can be planned. Agent cost scales with detours.
Hard caps are the obvious answer
Which is exactly what Simon Willison was after two days ago. He argued that hard budget caps should be the default, not an option buried in a settings page. The WSJ now supplies the number behind that argument: nine out of ten companies cannot say what they are about to spend.
Two tools against one problem, and they fit together. Anthropic has been showing the gateway budget in dollars since 2.1.284 instead of abstract units. That helps you look. What the WSJ describes needs more than a readout, though — it needs a limit that holds when nobody is looking. As long as the ceiling is an estimate, AI stays the one line in the budget where you start the month not knowing whether it will come in at $5,000 or $50,000.
Sources: