Some enterprise AI token budgets tripled in 2026 as firms struggle to measure returns

McKinsey finds only 6% of companies see measurable AI financial returns

The WSJ Technology Council Summit convened in New York this week against a backdrop of mounting enterprise AI spending. Executives on the panels said they remain bullish on AI’s promise to transform business while acknowledging that no one really had a handle on their 2026 token budgets, with some saying their initial planned budgets had as much as tripled this year. The gap between AI costs and the value the technology delivers was a central topic of the discussions.

“It’s hard to measure a lot of the ROI that happens,” said Arvind Jain, founder and chief executive of AI-powered enterprise search and work assistant platform Glean, speaking at the event.

Executives agreed that AI is delivering outsize value in some cases, but they found that in the majority of cases, it is wasting more money than it is worth. Often companies have a hard time distinguishing which is which. Better tools for observability are being developed but are still nascent, they said.

McKinsey’s Kate Smaje presented the consultancy’s latest research, which found that 80% of workers feel they are getting a productivity boost from AI. Only 6% of companies are getting real, measurable, financial ROI they can show to investors, according to that research. Smaje cautioned that individual task-level time savings do not necessarily translate into enterprise-level value.

“If I take 20 minutes on something and I multiply that out by my day, that doesn’t mean that somebody paid me 20% less or it doesn’t mean necessarily that I can get rid of 20% of people because it was all at the task level,” Smaje said.

“I think the reality is, no, most people are not getting value from it [AI]. But those that are, are incredibly concentrated,” said Smaje, McKinsey’s senior partner and global leader of technology and AI.

David Hsu, CEO of AI harness company Retool, which works with customers across the Fortune 500, said the results are “incredibly uneven” across uses. He estimated that about 90% of tokens companies spend are negative ROI, with the last 10% driving the majority of value. “We believe that about 90% of the tokens you’re spending are probably negative ROI. It’s just that last 10% of tokens that are really driving a lot of that ROI,” Hsu said. Hsu said he sees many companies getting a three-times to five-times return on their token spend overall, even though the distribution is uneven.

As an example of waste, Hsu said an unnamed company built an automatic AI-powered out-of-office responder that turned out to cost $10,000 a day. He described it as an extreme case but emblematic of a broader mismatch between spend and value that he said is pervasive in today’s enterprise landscape.

Jain said the first step to improving AI ROI is investing in dashboards that give extremely granular visibility into what everyone at the company is doing with AI. “When you allow people to build things, you give them immense power: the power to actually burn tokens and spend money,” he said. Jain described one instance in which a sales employee used AI to analyze lost sales opportunities, only for the analysis to cost more than the company would have recovered by signing the customer back. Better dashboards and observability platforms can help executives identify those use cases and cut back on them, Jain said.

Much of the wastage comes from using generative AI or agents for tasks they are not suited for, panelists said. Hsu said he worked with one company that was proud of having more AI agents than employees but was looking to optimize its costs. It became apparent, he added, that 95% of the problems the agents were solving were better suited to deterministic workflows — a shift that saved the company roughly $20 to $30 million.

Smaje said companies should match the AI model to the task rather than tapping the biggest frontier models for every use case. “You’re not trying to drive a Ferrari to the grocery store,” she said.

The most critical factor, Smaje added, is workforce upskilling and organizational change management. When she looks at the 6% of companies seeing meaningful AI value, she said, “they treat their AI transformation as a people transformation, not as a technology transformation.” Those companies, she said, match every dollar spent on technology with a dollar spent on change management, reskilling and upskilling. “All of that humanware becomes disproportionately important,” Smaje said.

Panelists emphasized the need to keep humans in the loop. Anthony Moisant, CIO and CSO of Indeed, said human judgment is “probably creating more lift in the business.” Smaje said the best AI deployments she has seen were 20% to 30% technology, with the rest contributed by humans. “Some of the best AI deployments that we’ve done, the technology has been 20, 30% of the answer,” she said. “And then it’s what the humans have put on top.” Galina Antova, CEO of AI cybersecurity company Kai, said the stakes are highest in cybersecurity. “It is up to us, the humans are in charge of the AI. It is not the other way around,” Antova said.

Marina F. Bellini, president of MGS & Digital Technologies at Mars, said companies cannot think about the human side and the tech side separately. “We cannot think about the human side and the tech side separately,” Bellini said. Bellini added that AI is helping break down long-established silos between business leaders and IT teams. “We spent years training IT professionals to learn the business language, and now we are spending years training the business leaders to speak the technology language. So it’s good we’re all going to speak the same language finally,” she said.

Additional summit takeaways included comments from XBOW founder and CEO Oege de Moor, who estimated that open-weight models lag frontier labs by only two to three months, a gap that drives down costs for both attackers and defenders. Moisant, of Indeed, added a cautionary note about leadership teams facing intense board and market pressure to move rapidly. “There’s just a rational exuberance to a degree where… movement means progress, like its frenetic energy,” Moisant said. “The caution that I try to embed in the organization is like, look, we all know that we can run really, really, really fast in the wrong direction.”

Work-Bench co-founder and general partner Jonathan Lehr said the enterprise buy-versus-build strategy has consolidated to “buy to operate and build to differentiate,” with off-the-shelf software covering 80% to 90% of capabilities and IT teams focusing on the last mile.

Doug Madory, head of internet analysis at cybersecurity firm Infoblox, described a recent incident in which a cloud region lost customer data because of an airstrike. “A cloud region losing customer data because of an airstrike — I don’t think that’s ever happened before,” Madory said.