The AI Spending Problem Isn't About Tokens.

Recent headlines about AI spending have focused on token consumption, model costs, and stories of organizations discovering unexpectedly large AI bills. One company reportedly accumulated roughly $500 million in AI expenses in a single month after usage expanded without meaningful constraints. Around the same time, reports emerged that Amazon shut down an internal leaderboard that rewarded employees for AI usage after workers began “tokenmaxxing,” optimizing for token consumption rather than business outcomes. Executives at other firms have similarly begun questioning whether rapidly growing AI expenditures are translating into measurable improvements in productivity, innovation, or customer value.

These stories are often framed as budgeting problems. In reality, they reveal something more fundamental.

Organizations are discovering that AI adoption and AI value are not interchangeable concepts, and many lack the governance mechanisms necessary to understand the difference.

Part of the challenge stems from how organizations measure success. During the early stages of adoption, usage metrics offer an easy way to demonstrate momentum. Leaders can track licenses issued, prompts submitted, tokens consumed, agents deployed, and hours spent interacting with AI systems. While these indicators provide evidence that employees are engaging with new tools, they reveal very little about whether those tools are creating meaningful value.

The rush toward AI adoption is understandable. Leaders face pressure from boards, investors, customers, employees, and competitors to demonstrate progress. Under those conditions, usage becomes an attractive proxy for success because it is visible and easily measured. Yet organizations that focus primarily on adoption often discover that adoption itself is not the outcome they actually care about. The objective is improved performance, stronger decision-making, greater resilience, new capabilities, or increased revenue. AI usage only matters to the extent that it contributes to those goals.

This dynamic reflects Goodhart’s Law: when a measure becomes a target, it often ceases to be a useful measure.

The tokenmaxxing phenomenon illustrates the problem. When employees are rewarded for AI usage, they naturally find ways to increase AI usage. The resulting activity may look impressive on a dashboard, but it does not necessarily improve customer outcomes, accelerate innovation, strengthen decision-making, or increase organizational performance. Organizations can find themselves celebrating adoption while struggling to explain why costs continue to rise faster than benefits.

The assumption underlying many AI adoption programs is straightforward: if employees use more AI, productivity should increase. In practice, the relationship is rarely that simple.

An employee can generate twenty reports instead of five without improving the quality of decisions those reports inform. An engineer can run multiple coding agents simultaneously without shipping more valuable software. A marketing team can generate hundreds of pieces of content without improving customer engagement or revenue.

The problem is not that AI lacks value. The problem is that organizations often mistake activity for outcomes.

Businesses have spent decades learning that website traffic is not revenue, training hours are not workforce capability, and lines of code are not software quality. AI introduces a new version of the same challenge. Tokens, prompts, and agent activity are useful operational metrics, but they are poor proxies for business value.

As AI spending shifts from fixed software licensing to usage-based consumption models, this distinction becomes increasingly important. Employees can launch agents, generate extensive outputs, run complex workflows, and consume significant computational resources without creating a proportional improvement in organizational performance.

The result is that many organizations now face a new question: not whether employees are using AI, but whether that usage is producing outcomes that justify the investment.

One of the most overlooked drivers of AI waste has little to do with model pricing and everything to do with implementation choices.

Organizations frequently focus on the cost of AI while overlooking the cost of automating the wrong work.

A workflow that costs $50,000 per month but eliminates $500,000 worth of repetitive labor may be an excellent investment. The relevant question is not the absolute cost of the system but the opportunity cost of maintaining the status quo. Conversely, even relatively inexpensive AI deployments can destroy value when they automate activities that were never meaningful constraints on organizational performance.

This distinction helps explain why some organizations struggle to demonstrate ROI despite widespread adoption. The issue is not necessarily that the technology is failing. In many cases, organizations are applying automation to tasks that depend heavily on judgment, trust, context, or accountability while underinvesting in opportunities where AI could significantly enhance human performance.

This article continues by examining why organizations are automating the wrong work, how leaders can distinguish between automation and augmentation, why agentic systems change the economics of AI deployment, and how governance becomes a capital allocation capability rather than a compliance exercise.