Did Nobody Forecast the AI Token Bill?
By Charlie Liu ·
Picture being the CFO of the firm that spent the last two years telling every client to go all in on AI, and then having to explain why your own AI bill got away from you. That's roughly where Yuval Atsmon found himself this year. He's McKinsey's CFO, at a firm that has been about as pro-AI as it gets.
So when he told Bloomberg's CFO Briefing this, it was hard to brush off.
"There is no way that I can afford for it to go up between now and next July the same way that it went up this year."
He wasn't calling for a pause though. "I wouldn't recommend anyone stop making AI investments at this moment," he said. What he was calling for was discipline instead of inertia, an actual forecast instead of a number that just happens to you.
AI token cost moves with usage, and usage moves with adoption, so the number doesn't sit still. Airia found some companies were three times over their entire 2026 token budget by April. A seat based license never swings that hard.
This is what I'd use to forecast it properly.
- Separate fixed and variable. Subscription fees are predictable. Token spend isn't, forecast it from actual usage, not a calendar.
- Build the forecast from a usage model. Requests times tokens per request times price per token gives you a real baseline.
- Revisit it every quarter, not once a year. Usage patterns shift too fast for an annual number to hold.
- Allocate the cost by team or feature. Someone needs to own the number, or nobody can explain a spike.
- Forecast in cost per outcome, not total spend. Cost per query or per report, not just a top line number.
If McKinsey's own CFO needed a real forecast instead of a guess, the rest of us probably do too.
Sources
Last updated