Spending more tokens doesn’t always translate to value. A report released by OpenAI last week analyzed over 1,500 organizations and found that revenue per employee wasn’t meaningfully associated with an employee’s token output.

Yet while companies struggle to tie token consumption to ROI, AI spend continues to increase. According to the Ramp AI Index , which uses spend data from over 70,000 firms using its corporate card and bill payment platform, the top 1% of businesses spent a median of $7,400 per employee on AI, while the top 10% spent $650 in July 2026. For comparison, the top 1% spent $2,590 in January 2026, whereas the top 10% spent $281.23.

Goldman Sachs also estimates that token consumption will multiply 24-fold to 120 quadrillion tokens per month between 2026 and 2030, as consumers and enterprises adopt AI agents, while semiconductor providers are expected to deliver cost reductions of 60%-70% per year per token for inference.

As spending on tokens increases, organizations will face increased pressure to demonstrate value. However, while measuring the ROI of AI adoption remains difficult, particularly within organizations that are ‘ tokenmaxxing ,’ tokens provide an imperfect metric to measure AI use.

If left unchecked, token spend can add up quickly in the enterprise. Back in May, Uber’s COO made headlines after an interview in which he claimed the company had blown through its AI budget in just four months. Similarly, in July, The Financial Times reported that Amazon saw cost overruns of $1.8 million following a failed Claude Sonnet deployment.

Recently, OpenAI chairman Bret Taylor predicted that companies will stop worrying about AI tokens, but G2 research finds that 80% of software buyers now provide developers or technical teams with a token or LLM usage budget, indicating AI costs are being budgeted and tracked like any other line item.

Tim Sanders, chief innovation officer of software comparison platform G2, told me in a video interview in July that the company has spent over $1.27 million on AI tokens in 2026, the equivalent of 970 billion tokens, with Claude Code and Cowork seeing increased adoption among employees, accounting for 17% of token spend in January, but 79% in July. “I remember when companies barely offered employees T1 access,” Sanders said. “I worked at Broadcast.com and Yahoo, where there was no Wifi. There was a debate over the return on investment of putting T1 servers into companies and giving employees broadband access, but soon they realized it was a utility. If you didn’t believe it was a utility, just turn off the servers and watch everybody grind to a halt. Now companies blindly purchase Wifi and laptops for their employees. They don’t run ROI because they realize they’re foundational.”

Sanders says that G2 looks at AI like electricity, but tries to be “prudent” with spend, looking for time savings, cost savings and the ability to increase income. Instead of tokenmaxxing, or optimizing for total tokens used, G2 manages consumption by measuring token efficiency among employees. More specifically, the company maintains a company-wide internal AI Efficiency Index that identifies the team members getting the most out of each token. The index’s formula is Tokens / Dollar * In(1 + Tokens / 1,000,000).

The company also shared data about the efficiency of some of its superusers, claiming that Mike Wheeler, G2 cofounder and CTO, maintains a token efficiency of 9.54 million tokens at $1.09 tokens per dollar, and Dan Knox, G2’s VP of Engineering, maintains a token efficiency of 9.55 million tokens at $1.34 million tokens per dollar in July.

In terms of measuring value, Sanders emphasizes output quality rather than hours saved via automation. “We’re not just measuring how much time we saved; we’re also measuring our velocity in delivering outcomes against our key objectives."

Spending tokens is meaningless if it doesn’t translate to measurable value, but organizations that are selective about what workflows they automate can increase the chance that spending on agents is impactful. For example, Zapier’s AI Workflow Index , released in July, analysed AI usage from over 1,500 organizations, and found that among the leading adopters, AI shows up in just 18% of workflow steps, while 82% of steps run on code and logic.

“The sort of savvy folks are figuring out when to use it and how to use it best. AI is very, very effective, but it is not a panacea for all problems,” Wade Foster, co-founder and CEO of Zapier, told me in a video interview. Illustrating this point, Zapier finds that workflows that use AI only for judgement calls cost 71% less to run than workflows that route every step through a model.

Foster says AI delivers the greatest value in four distinct roles: as a communicator that writes content, a clerk that extracts structured information from unstructured documents, an analyst that makes judgement-based decisions, and a coordinator that assigns work across teams. Everything else, he says, is often better handled through conventional automation.

When asked how many companies are using agents effectively, Foster said, “most are probably not,” pointing out the trend of tokenmaxxing, which he defines as “burning lots of tokens and doing it relatively inefficiently.” He adds that this approach is seeing “blowback” from CFOs who are questioning the returns gained from token spend.

That being said, Foster says internally that non-technical employees can spend hundreds of dollars per month while most developers at Zapier spend thousands of dollars per month, with top users spending as much as $30,000 per month. Those high-spending developers are generally working on Greenfield projects, where AI coding agents have clean feedback loops and iterate on bug fixes and new functionality.

Many organizations use token caps to limit token spend, but Angel Lange, CFO of Agiloft, an AI contract management and contract lifecycle management platform, suggests this can be an ineffective control.

“We’ve learned that managing AI costs is about limiting token usage and understanding where the value is being created,” Lange told me via email, adding that 74% of users weren’t hitting AI usage caps, which indicated to the company that broad restrictions weren’t addressing the real drivers of spend.

“Instead, we shifted our focus to visibility: understanding which workflows consume AI resources and whether they’re delivering measurable business outcomes,” Lange said. ”The question isn’t, ‘how many tokens did we use?’ It’s, ‘what did those tokens enable? If they’re accelerating delivery, improving quality and freeing employees to focus on higher-value work, then token spend is an investment in productivity, not simply a cost to minimize,” Lange said.

It’s worth noting that Agiloft has been using AI internally as part of its workflows, closely managing token usage, and reports a 67% increase in productivity among developers attributable to the use of AI tools.

Measuring token use provides a rudimentary way to assess AI adoption, but connecting consumption to value can be more challenging. Developing visibility over spend plus time saved, outcome velocity and employee productivity can make the connection between AI spending and business more clearer.