The Information surveyed 107 executives and found 60 percent said their AI cost trajectory is unpredictable, and only 33 percent actually control token costs day to day

The Information reported on 21 September 2026, in a survey of 107 senior leaders and technology professionals, that 60 percent describe the cost trajectory of their AI initiatives as unpredictable, at the same moment that CFOs and boards are asking for a hard number. The outlet also reports that 52 percent of respondents say they understand how token pricing works, while only 33 percent say they actually control those costs in day-to-day practice. The Information ran the survey among its own readership and publishes the paywalled article behind a subscription.

The gap between "we understand it" and "we manage it"

The pair of numbers to read carefully is 52 and 33. Half of respondents can explain what a token is and why the API bill scales the way it does, and a third of respondents are running their workloads in a way that keeps the bill in a chosen range. The 19-point gap is roughly the width of the operational discipline needed to close it: caching, prompt trimming, choosing a cheaper model for a step that does not need the frontier one, batching similar requests, and cutting off agents that keep looping past a token budget. None of those levers are secret. Each is a specific engineering decision that can be assigned, measured and reviewed. The survey suggests that in most companies they are simply not.

Why the unpredictability number should worry a technical leader

A budget line that comes back with a different number every month is the fastest way to lose executive support for an AI programme. Traditional enterprise software carries a seat-count and an annual price, so the finance team knows what the year will cost by the second week of January. A token-priced API bill scales with usage, and usage scales with product decisions the finance team never sees, so the December bill often bears no resemblance to the July one. The Information's number, 60 percent unpredictable, is the industry-wide version of a conversation that a lot of engineering leaders have already had with their CFO in private.

What a team can do this week

For a team paying an AI bill that grew this quarter, the survey does two useful things at once. It puts a number on the industry pattern, which is how to argue for engineering time on token discipline in front of a finance team that has heard the word "AI" too many times to react to it. And it names the specific gap that time should close: the operational plumbing that turns a variable-cost API into something the finance function can plan against.

That means putting the token count and dollar cost of every production call into the same dashboard the on-call engineer reads, alerting on a per-feature per-user cost that exceeds a fixed budget, and running a monthly review of the ten most expensive endpoints with the same seriousness as a monthly cloud bill review. None of that is new, and by The Information's numbers, most of the market has still not done it.

Source

The Information, The Real Cost of AI: A Survey of the Unpredictable Token Economics, 21 September 2026. Paywalled.