Employers cut back on AI as costs soar

"Never in the history of the world that I know of has a company gone around and handed every employee a credit card and said, ‘Spend freely and we're going to reward you based on how much you run up on the credit card.’

“But this is the token team logic."

So says David Kryscynski, professor of strategy at the School of Management and Labor Relations at Rutgers University, in discussing the growing concern around AI costs and a lack of obvious ROI.

Many companies that spent the last two years encouraging employees to try AI tools with few guardrails are now staring at budget overruns they didn't anticipate — and struggling to show what they gained from it.

Uber Technologies, for example, exhausted its entire planned 2026 AI coding budget within the first four months of the year due to soaring token consumption — the basic unit AI companies use to bill customers, representing the volume of text processed — and has since capped monthly employee spending at $1,500 per tool.

The chief operating officer said that it’s highly ambiguous whether increased AI token spend translates to tangible outcomes for the business, saying it’s “getting harder to justify” the costs.

In some cases, the cost of tokens is exceeding the cost of the employee within a month or two of use, just because of excessive usage, according to analyst Jack Gold in a Yahoo!Finance article. Microsoft's Experiences + Devices division ordered a cutoff of third-party Claude Code licenses by June 30, 2026 while Meta's internal "Claudeonomics" leaderboard tracked 85,000 employees burning more than 60 trillion tokens in 30 days before being shut down due to waste.

Workers ‘rewarded for consumption’

"What we can see really, really, really clearly is the costs. We can see the money going out to pay for tokens. That's obvious. The return, however, is much less measurable and much less clear,” says Kryscynski.

"We're rewarding people for consumption."

The dynamic has a familiar shape to anyone who has watched performance metrics go wrong before, he says.

“The incentive problem in organizations is that we measure what we can measure easily and we create KPIs around things that we can measure easily. And what becomes measured becomes rewarded — and measures become incentives in and of themselves.”

Token consumption has become a default proxy for measuring AI engagement. But it tells an incomplete story, according to Carolyn Hamer, managing director and global human capital practice leader at Deloitte.

"It tells us about how much people are using the tools and I think that's a dimension of it. But it doesn't really get at impact,” she says.

“When you're talking about ROI in token consumption as an example, this is looking at ‘How much AI did we use?’ versus ‘What value did it create? And I think… that is where we want to be moving… How are we creating value that didn't exist before, how is work changing, how are we improving outcomes for both the business as well as our employees in ways that we didn't foresee or couldn't have actually done in the past?”

Early days for ROI

Traditionally, ROI could be about per unit production cost, for example, but that doesn’t apply to AI, says Kryscynski. Another consideration is when people learn new tools and skills they may unlock unrealized potential.

“There's multiple pathways for a potential return, but it's very hard to anticipate those in advance because we don't have models for it and we don't know what it's going to look like,” he says.

“We're still in this Wild West kind of period where the benefits are potentially still long delayed and haven't yet been realized.”

Hamer agrees that there’s a latency problem that complicates any short-term ROI calculation. Some of what companies are spending now is foundation-setting — getting people comfortable with tools, building new ways of working and leading.

"It's hard to see that value translated into both hard measures but even soft measures in the immediate term… it's early to be seeing that impact, but I do think there's impact that is intangible in terms of the foundation that's being set,” she says.

“Our ability to get to hard measures — especially impact value, workforce productivity, workforce engagement, those things — is going to be harder to determine at this stage.”

Should AI use be more targeted?

Meta, which earlier this year encouraged employees to use as many tokens as possible as a measure of productivity, had second thoughts, according to yahoo!finance. Chief technology officer Andrew Bosworth wrote in a memo to staff: "Nobody should be using AI tools just for the sake of using them."

As organizations look to rein in spending, many are pivoting from open-ended experimentation to more focused AI use. And it makes sense to “calm down” the excessive token use, says Kryscynski — but he warns that overcorrecting carries its own risks.

"When we say targeted then, well, what defines targeted? What defines what we use AI for and what we don't use AI for? And what defines what's an allowable use of tokens and not allowable use of tokens?” he says.

“When we start to do that, we start to step back and say, ‘Let's take our old model of work to define what's appropriate use of AI’ — and that might be inappropriately limiting the potential of AI to actually add value in your organization."

The right approach, he says, depends on the type of organization. A cost-focused company can reasonably narrow its AI use to repetitive, error-prone or unpleasant tasks where the efficiency gains are clear. But for organizations at the forefront of discovery, targeted use becomes much more difficult to think about “because if we’re trying to push the boundaries of what we can do, then targeted seems to be almost self-limiting to the objective of the organization."

"It's definitely a double-edged sword,” says Kryscynski, admitting he’s not quite clear on the solution.

Old habits, new tech

The turbulence isn't entirely unfamiliar. Kryscynski notes that every major technology adoption follows a similar curve: initial costs rise as organizations navigate the learning period, efficiency dips before improving, and the full value takes time to emerge.

The difference with AI, he argues, is structural. When companies gave employees personal computers, the marginal cost of each additional use was essentially zero. AI tokens work differently.

"If you build an army of 1,000 AI agents that each then build their own armies of 100 or 1,000 AI agents doing these little tasks — that may or may not ultimately have value — then now we're significantly adding to the marginal cost of use of this new technology."

The result is a fundamentally different cost structure than organizations have encountered before.

"We're decentralizing decision-making over the costs," says Kryscynski.

False choice: AI vs people

Some employers have framed AI investment as an alternative to headcount — but research on similar technological shifts suggests the trade-off isn't straightforward, says Kryscynski. Studies on robotics adoption have found that as firms invest in advanced technology, they often end up needing more people, and higher-skilled ones at that, to operate it effectively.

"I suspect that as organizations engage AI more and more, what they're going to find is that they need a different set of skills and new tools and new people with different types of skills to be able to work in a new workflow in a new organization," he says.

Hamer agrees the framing of AI versus humans is the wrong one.

"What we want to move to is really that human powered by machines — so, how do we get the humans and machines working collaboratively such that… the fabric of the workforce is just natural and organic and those handoffs between human and AI feel seamless?"

HR's role: from access to accountability

A year ago, the dominant question in HR was whether employees had access to AI tools. That has changed and now the question is what to do with that access.

"HR's role is to help be really clear on what are the expectations around AI usage, how does that translate into performance and how are we starting to think about changing our reward and recognition approaches to account for those expectations?" says Hamer.

"I don't think we're quite there yet, but I think that's going to be the next step."

For HR teams, it should be about measuring how AI is improving quality, decision speed or employee experience, says Hamer — “things that are really going to get at the fabric of how individuals are managing and succeeding at work.”

Kryscynski agrees the prescription is to shift the focus from measuring consumption to evaluating outcomes and creating value.

“Rather than incentivizing and rewarding token usage, let's think about the human judgment around quality work and let's ramp up our human judgment and evaluation of the work that's being done.

"We're not going to reward volume of tokens. We're going to reward incidents of creativity. We're going to reward experiments that failed but that we learned from. We're going to reward instances of AI use that lead to productivity gains or cost decreases."

It’s about a different type of management, and HR is well placed to lead that shift and “add a ton of value” to this conversation, according to Kryscynski.

"What we need to do is think about ‘What is the work that matters in this work unit?’… and then rather than incentivizing and rewarding token usage, let's think about the human judgment around quality work."