Recently, Uber's COO said AI costs are getting "harder to justify." Microsoft canceled most of its Claude Code licenses, citing spend. And according to Axios, one organization burned through half a billion dollars in a single month after failing to put usage limits on employee AI licenses.

There's a term for the pattern that drove these situations: tokenmaxxing. The organizational push to use as much AI as possible, as fast as possible. It emerged from two years of competitive anxiety, where the pressure to adopt AI outpaced serious consideration of long-term ROI. In the absence of true metrics, organizations created usage leaderboards, which people quickly learned to game, and usage soon became a proxy for value.

The bill is starting to come due. And in most organizations, the CFO is the one holding it.

The cost structure hasn't caught up to the ambition

Most enterprises still don't have the financial models and discipline to accurately measure what AI costs them. They can track what they're paying for AI licenses, but few organizations have a reliable framework for attributing token consumption, compute, and spend to any real business outcome.

For the past few years, many organizations have been chasing the “AI dream.” It was seen as a technology initiative, rather than a business transformation, and nobody was asking whether the return on AI investments was showing up on the P&L.

According to research from the IBM Institute for Business Value, seventy-nine percent of executives expect AI to drive significant revenue by 2030. Only 24 percent know where it's going to come from. If you can't define the outcome, you can't measure the return. And you can't defend the spend to a board.

What disciplined AI adoption looks like

The fix isn't to cut, it's to focus. When AI is layered on top of existing workflows as a general-purpose productivity tool, token spend has no anchor. When it's embedded in a specific process tied to a specific outcome, the economics change.

First, every AI use case needs to connect to a specific workflow and a measurable outcome, with returns tracked in three-to-six-month increments. Like any other capital allocation, if it can’t deliver a 2.5x to 3x return, whether through time saved, better customer and employee experiences, or new revenue, pass on it.

Second, build the orchestration layer. There's been no shortage of conversation about smaller, more efficient models. But that's only one variable and it won’t drive returns on its own. Business leaders need to architect AI as one cohesive system, and that starts with being able to identify and monitor which model is best suited for each task or problem, then ensuring every part of the system routes work accordingly.

IBM Bob, for example, orchestrates across multiple models automatically, optimizing for cost, quality, and performance. IBM Consulting Advantage, an AI-powered delivery platform, includes a model router that directs each task to the appropriate model seamlessly.

Third, this can't be owned by one function. Token spend is a financial, architectural, workforce, and strategic problem simultaneously. Research from the IBM Institute for Business Value shows that 68% of executives worry their AI efforts will fail due to lack of integration with core business activities. Without C-suite alignment, AI stays siloed rather than embedded across the business, and the ROI never materializes.

The euphoria phase is ending, but it’s giving way to something more durable. What comes next is harder and more important: building the financial discipline to treat AI like any other capital decision. As scrutiny increases, organizations will need to show not just where AI is being used, but what, precisely, it’s worth.

Click here to see this post by Neil Dhar on LinkedIn.