By Belle Lin

If there's one thing companies know how to do, it's slam on the brakes when spending starts to run hot. But artificial intelligence is testing that instinct. AI costs are climbing, though most tech leaders remain convinced it can eventually deliver real returns — just not at any price.

Chief information officers told The Wall Street Journal Leadership Institute they are deploying a number of strategies, including tried-and-true techniques sharpened during the rise of cloud computing — and the need to manage ballooning cloud costs — to keep their AI costs under control.

"With AI, you're putting the credit card in the hands of the end user. If you have no control over that, or if the end user is not educated enough, they're going to run up that tab," said Chris Reed, a senior director of IT finance at online travel company Priceline.

Unlike in previous tech cycles, corporate adoption of AI rests on all employees — not just developers — picking up on the technology. AI is increasingly being billed by usage, and the price of tokens, the basic unit of AI computing, has been volatile. That all translates to higher costs for AI.

At Priceline, dashboards track token usage, with monthly reports delivered to the chief financial officer and chief technology officer, said Reed. High token usage "prompts a conversation" with the employee to figure out how they've been using AI — and limits can be malleable if they're working on a revenue-generating initiative, Reed said.

Adding to the cost pressure is the shift from prompt-based chatbots to always-on autonomous AI agents, which consume vastly more tokens. And with larger, more sophisticated AI models, those costs are expected to climb sharply.

"It will be orders of magnitude higher than what we spend today," said Greg Meyers, chief digital and technology officer of Bristol-Myers Squibb, adding that he expects "exponential" costs associated with AI agents as AI usage hits an inflection point.

Meyers said he has already prepared the biopharmaceutical company's management team, including its CFO and board, to expect "pretty high token consumption."

But it isn't all bad news: "If you factor in what we believe is the payoff here, we believe that it's actually a pretty positive [return on investment]," he said.

Breaking out the cloud toolbox

Some corporate tech leaders are dipping into their toolbox for cost-control strategies honed during the rise of cloud computing.

Kathy Kay, CIO of Principal Financial Group, said the financial-services firm is "putting governance and optimization practices in place, similar to what companies have done with cloud, to manage costs as we scale." For instance, Principal is focused on using the right AI model for the right task, so that "higher usage doesn't necessarily translate into higher costs," she said.

"Given how quickly pricing and capabilities are evolving, we're designing for flexibility so we can adapt over time and continue to deploy AI efficiently," Kay added.

FinOps — a blend of finance, engineering and product — emerged from the cloud-adoption boom of the 2010s as an approach to managing and maximizing the business value of technology and cloud spending.

Ravi Soin, CIO and chief information security officer of Smartsheet, said the company's FinOps team is responsible for tracking its overall AI spend. "Someone really needs to own the budget," he said. The software company has set automated alerts so that employees know they're about to hit their token limits.

"We have user dashboards available to the entire company, by department, by manager, so you have real-time visibility on how often and what your costs are, so it isn't a surprise at the end of the month," he said.

Other companies are seeking to hire outside experts with FinOps experience. CVS Health is hiring for an AI Ops Engineering executive director with expertise in FinOps — that includes what the company described as "GPU cost governance," and "cost reduction," according to a job posting.

Agents, agents everywhere

Enterprises already are using more AI than ever before, with many wrangling more AI agents than they can keep track of.

Compared with asking a chatbot a question, asking an agent to complete a task can require 50 times as much computing power, according to Jim Schneider, a senior equity research analyst at Goldman Sachs. Goldman predicts that AI agents will increase AI token consumption by 24 times over the next four years, and business AI agents will increase token consumption by 55 times by 2040.

Model providers OpenAI and Anthropic have said the costs of their tokens are going down, and both have considered drastic price cuts.

Even with less expensive tokens, however, agents are consuming more of them as they interact with other agents and work over long periods of time. While model prices fell roughly 50% from December 2024 to December 2025, tokens consumed grew 4.5 times in the same window, according to research from Bain and Co.

Lowering the AI bill

The challenge isn't just controlling what AI costs — it's knowing what it's worth, tech leaders said.

Atilla Tinic, CIO of Qualcomm, said the semiconductor company has used a few different techniques to keep AI costs in check, including placing caps on how many tokens certain teams can use. "Some of the engineering functions are obviously large consumers of tokens," he said.

Qualcomm has also instituted a "show back" to various corporate departments — a way to show them "dollars tied to the token, so that people understand what their spend does," Tinic said.

OpenText CIO and Chief Digital Officer Shannon Bell said the show back or "chargeback" method can lower an organization's token costs by 20% to 30%. "We want our development leaders to be accountable and take ownership for the spend, and the outcomes that they get as a result," she said.

But it's hard to capture all the nuances of AI usage.

"High AI usage isn't necessarily a good or bad thing. It depends on the business outcome that's attached — that's the most difficult part to quantify," said Priceline's Reed.

That uncertainty is pushing companies toward another tactic: paying less per task. Rather than run everything on large, expensive models, some are swapping in smaller, older or open-source models. Running those models on Qualcomm's own hardware saves the company even more, Tinic said.

Seemantini Godbole, chief digital and information officer of Lowe's, said the company is putting guidelines and mechanisms in place to avoid "token wastage," including using smaller and open-source models.

"Token usage is a really good thing, as long as it's in the interest of meeting a business objective," Godbole said. "We also need to make sure that we are not unnecessarily spending capital where we don't need to."

Write to Belle Lin at belle.lin@wsj.com