Companies are pouring AI into coding, finance, marketing, and customer workflows, but many finance chiefs still cannot answer a basic question: how much is this thing costing us? A KPMG survey, not yet published, suggests the answer is often ”not much better than a guess”, and that is becoming a serious budgeting problem as more vendors move to usage-based pricing. The issue is especially acute for AI token usage, where costs can rise quickly and unpredictably.
Only 26% of companies say they have a full view of AI spending, while 50% have only partial visibility. Another 22% learn the size of the bill after invoices arrive, or lack enough transparency to know at all. That is a familiar story from the cloud era, except now the meter is running in tokens, cloud compute, and AI agents that can chew through budget far faster than a software license ever did.
Why AI token usage is breaking old budgeting habits
The shift is simple and annoying: instead of paying a fixed fee for software, many companies now pay for actual usage. OpenAI and Anthropic use this model, and so do enterprise vendors such as Microsoft and Salesforce. For providers, it lines up revenue with compute costs. For customers, it pushes risk downstream and makes forecasting much harder.
KPMG’s global AI lead, Steve Chase, said some companies have burned through a full year of token and cloud budgets in just a few months, and one client saw token consumption jump sixfold. That kind of spike is exactly why finance teams are now treating AI usage like a line item that needs constant supervision, not a shiny pilot that can be left to the tech team.
Life360, Affirm and Reckitt are already tightening control
Some companies are moving quickly to get a grip. Life360 is building tools to reduce token use and redesigning its AI-agent architecture, even though it still lacks full real-time spend monitoring. Affirm, meanwhile, saw token consumption rise almost instantly after expanding AI agents for coding, and now tracks usage close to real time while reporting weekly to leadership.
- Life360 is trimming token consumption and reworking AI agents.
- Affirm monitors usage almost in real time and reports weekly to executives.
- Reckitt found some employees drifted back to old workflows, cutting tool usage after an early launch of 12 marketing AI solutions.
Reckitt’s experience is a useful reminder that AI spend is not just about raw demand; adoption can wobble, data quality can slow rollouts, and promised payback can stretch. The company still plans to keep investing, but it has already learned that a rollout can look impressive on launch day and far less tidy a few weeks later.
The next AI battle is cost control, not just adoption
Some companies are responding the way enterprises always do when a new bill starts getting ugly: by rationing access. Corning has narrowed the AI tools available to staff and is concentrating budget on a smaller set of bigger projects, while still allowing experimentation. Amer Sports is taking a slower route, testing AI for invoice handling, financial close and reconciliation, but deliberately pacing rollout to avoid runaway token and agent costs.
That caution feels overdue. The first wave of corporate AI was about proving usefulness; the next will be about proving that usage can be controlled without strangling the gains. If finance teams can finally track tokens with the same discipline once reserved for cloud spend, the loudest AI winners may be the ones that learn to say no.