Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

How much is your AI token bill?

A sense of AI urgency has led to unbounded spending and vague metrics.

[Photo: Getty Images]

Profile photo of Jeremy Sack

President, Insights · Material

A finance lead, an IT director, and a line-of-business executive sit around a conference table staring at a monthly cloud invoice that is several orders of magnitude higher than expected.

The culprit? AI token usage.

They wonder how this happened.

IT thought the business unit was tracking usage, the business unit assumed finance had budgeted for it, and finance assumed it was just part of standard tech overhead. Nobody had a clear view of how AI tokens were being allocated or spent, which made the problem worse.

During early adoption, leaders rushed to get something—anything—deployed. They’d figure out the details later. But when later arrived, the sticker shock wasn’t just a finance problem. It forced a long-overdue conversation about whether AI investments are delivering business value.

How we got here

For the past several years, enterprise AI deployment was unduly influenced by a sense of urgency. Companies launched pilot after pilot because boardrooms demanded action. Falling behind was unacceptable, even if measurable value or ROI were undefined.

That sense of AI urgency led to upbound spending and vague metrics. Now, reality is setting in. Quantifiable token costs and ROI demands are leading to questions about why and how the company should deploy AI. For months, AI spending remained abstract while metrics for AI success were either never established or they were completely qualitative, often measured by statements like, “It feels faster” or “The team likes it.” Now, the bill has arrived, and it is very quantitative.

Some high-profile examples of out-of-control token usage include Uber depleting its AI budget in four months; OpenClaw spending $1.3 million in 30 days; and an unnamed company receiving a $500 million AI bill for excessive Claude use.

Token costs are the first hard number most organizations have had to reckon with in their AI adoption journey. Unlike early social media or SaaS experimentation budgets—which were often small enough to be swept under the rug—AI spending is significant enough to land squarely on the CFO’s desk.

The scale of the spend has been driven by a pair of factors.

1. The ownership vacuum

Who owns the AI budget? Right now, the answer is often messy.

  • IT controls the infrastructure and API keys.
  • Line of Businesscontrols the workflows and prompt usage.
  • Finance holds the checkbook.

With three partial owners and little accountability, costs can spiral unchecked.

2. The token consumption trap

A lot of enterprises incentivized their employees to adopt AI; but they did so without a value- or solution-focused strategy. They measured employee productivity in tokens generated, not by the quality of outcomes.

To climb productivity leaderboards, employees practiced token-maxing—writing massive prompts for simple (or unnecessary) answers, burning through tokens and budgets for irrelevant outputs of little or no value. The return on AI investment in these cases is negligible—and the excessively long and pointless prompts can harm AI model performance.

Disciplined AI spending looks like this

The good news? The era of unchecked experimentation is giving way to a more mature phase of enterprise tech. Token-maxing is being replaced with value-maxing.

Instead of asking, “How can we use an LLM here?” companies are beginning to ask, “What specific challenge are we solving for or what value are we hoping to create? Do we need an LLM to accomplish this?” Disciplined companies are also learning that they don’t always need the biggest, flashiest, and most expensive models and are more carefully matching the model to the task. And, instead of measuring AI success by adoption and deployment, they’re looking for efficiency, accuracy, and ROI.

The late-adopter advantage

If your company is just now scaling its AI initiatives, you avoided the early adopter tax. You can learn directly from the architectural and financial mistakes of the first wave, building cost-governance into your infrastructure from day one.

The truth is that not every workflow needs premium compute; most use cases don’t need the most expensive model or infrastructure. It’s important to avoid the token consumption trap. Use metrics that measure output quality against token usage. Have humans evaluate the usefulness of outputs and measure the acceptance rate of AI-generated code.

By finding ways to measure business value per dollar spent on AI compute, you’ll be able to make smarter decisions about what projects need AI investments and which ones are better left to other methods, including human intelligence.

Before your next AI initiative, ask these five questions:

1. What prioritized business problem are we looking to solve with AI?

2. Who’s accountable for planning, allocating, and monitoring AI resources?

3. What are the specific efficiency or revenue metrics that AI resources/workflows are tied to?

4. What do our unit economics look like if usage increases by the amount we’re allocating for next quarter? What if usage spikes beyond that?

5. Are we using a top-tier flagship model for tasks a smaller, localized model could handle at 1/10th the cost?

The bottom line

When the dust settles, the winners of the AI boom won’t be the companies that spent the most money on tokens or generated the most output. They’ll be the organizations that spent intentionally, where there were opportunities for creating real business value, and could prove the return on investment.

So, before your next invoice arrives: Do you know what your AI usage is costing you? More importantly, do you know what it’s worth?

Jeremy Sack is president of Material.