The AI Cost Paradox: Some Firms Save Millions Even As Token Costs Surge Across Tech

At cloud communications company 8x8, employees rely on Anthropic’s Claude AI chatbot for a wide range of work: drafting routine emails, analyzing batches of customer feedback, and writing new code. Yet as the company’s reliance on the tool grows, it has not sparked stress for 8x8’s finance team. While many major Silicon Valley players including Meta, Uber, and Salesforce have publicly raised alarms over generative AI’s skyrocketing costs and rolled out usage limits in some cases, 8x8 says it remains solidly profitable from its AI adoption.

Over the past 18 months, the company estimates it has cut roughly $5 million in annual costs by canceling subscriptions to dozens of legacy software and educational tools. Many of these tools became redundant, in large part because Claude already delivered similar capabilities, and 8x8’s total annual spending on Claude is “well below” that $5 million savings figure, according to Joel Neeb, 8x8’s chief transformation and business operations officer.

Neeb expects savings and AI costs will eventually balance out as 8x8 encourages more employees to adopt AI and integrates the technology into more complex workflows. But for now, the gap between what the company saves and spends remains wide enough that it “makes my chief financial officer happy,” he told WIRED. He declined to share exact total spending on generative AI.

As companies collectively pour hundreds of millions of dollars into AI tools for coding, marketing, and customer service, a new core focus has emerged across the tech industry: “tokenomics,” or the work of managing the soaring cost of AI usage. (Tokens are the unit AI models use to measure how much content they analyze and generate.)

Last month, Royal Bank of Canada’s CEO disclosed that the bank’s AI token usage surged 500% over just six months. At Cisco, one-third of employees use an internal AI chatbot daily, so “the token usage is getting pretty, pretty crazy,” CEO Chuck Robbins said during an earnings call. At analytics software developer Amplitude, some top engineers are “spending thousands of dollars a month or more on tokens,” according to CEO Spenser Skates. Box CEO Aaron Levine noted that “the token budgeting conversation has absolutely taken over as one of the most important” and “heated” topics for executive teams today.

A WIRED review of earnings call transcripts from data provider AlphaStreet found that roughly 300 companies addressed questions or concerns about AI tokens during earnings calls or public discussions with analysts in April and May of this year. While that makes up a small fraction of the thousands of calls held in that window, it is a major jump from the same period a year earlier, when just 93 companies mentioned “token” in any context.

Some company executives say they are building or shopping for new systems to track token usage and automatically select the lowest-priced model that can complete a given prompt. Others say they are still working to strike a balance between hiring new staff and increasing token budgets to hit their business goals.

Enterprise software has never been cheap, but the latest generation of AI tools has created unusual strain for C-suite leaders for several reasons. AI pricing is constantly fluctuating, new more powerful—and more expensive—models launch every month, and rolling out company-wide adoption of new work processes has proven challenging. Productivity gains from AI in one team can even create new bottlenecks for other teams that have not kept up with adoption.

20 Percent of Salary Budgeted for AI

That said, some companies are still encouraging employees to use AI freely without worrying about the bill. In April, Baseball Lifestyle 101, a Long Island, New York-based apparel brand on track to generate $250 million in sales this year, told roughly 50 of its top managers they could spend the equivalent of 20% of their annual salary on AI tokens each month.

Co-founder and chief strategy officer Bill Rom told WIRED total monthly AI costs will likely exceed $100,000 by the end of the year, but the investment has already paid off. Claude recently helped the brand land a $1 million order by flagging that a retail partner was running low on specific sizes of its popular ice-cream-patterned shorts. “That’s a day and a half of work that can now happen in an hour or two that might make me eight figures of additional revenue over 12 months,” Rom says.

The AI chatbot also helps draft financial reports and plan photoshoots, allowing the company to hire fewer junior staffers and redirect that payroll to other growth investments. Rom says it is more important to “inspire people how to use AI” before setting strict cost rules for the technology.

At 8x8, which builds communications platforms for sales and customer service teams, all roughly 1,800 full-time employees can access a public dashboard that tracks their own and their colleagues’ Claude usage, designed to ensure no one is left behind on the company’s AI journey. “It's not punitive in the least; it's really just so that we all stay tightly packed in this journey,” Neeb says. As of May, product and customer success teams were among the heaviest users, while sales and finance teams used the tool the least.

Neeb says 8x8 may eventually put caps on how much staff can use Claude. He first discussed the idea recently with the CFO, amid growing internal use of Anthropic’s new Claude Opus 4.8 model, which launched last month and costs nearly 1.7 times more than the Anthropic model released in February. Though no decisions have been made, future access to Opus may require employees to prove older cheaper models cannot complete the work, Neeb says. “Can we downgrade the model a little bit and still get the same outcome?”

Even with that possible change, Neeb says 8x8 is not pulling back on generative AI adoption at all. Customer satisfaction and loyalty scores have trended upward, and 8x8 has posted four consecutive quarters of revenue growth, in part because AI-generated analysis speeds up sales team work. While attributing those positive trends solely to AI is difficult, Neeb suspects a clear connection. “It really is the rising tide that floats all boats when you do this right,” he says.

“Go Faster”: Holding Employees Accountable for AI Adoption

About two years ago, 8x8 began exploring generative AI by giving all employees training and access to OpenAI’s ChatGPT and Google’s Gemini. Later, a small group of early users got access to Claude, Neeb says, and it became the company-wide standard over the past year.

Management monitors usage and has warned employees that refusing to build AI fluency will lead to consequences. “If you're not using AI in some capacity for your role, then you're missing the opportunity to go faster and get better answers more effectively than your peers,” Neeb says.

Other tech firms have issued similar directives with mixed results. At companies like Amazon and Meta, workers have reported using AI only because they feel forced to, or slacking off because AI frees up extra time, both of which critics call wasteful. Neeb argues patience and clear accountability measures are needed to orient employees to AI, and he does not want workers taking longer lunches or checking out while AI speeds up their work.

Neeb has also pushed for more adoption from what he calls “laggards in this journey”: sales and finance teams, which together make up 28% of 8x8’s employees but account for just 15% of the company’s total token consumption. He hopes a recent AI hackathon for the finance team will encourage the department to automate its extensive manual processes, such as customer collections and generating quarterly accounting reports.

Neeb has seen firsthand how Claude can help 8x8 work more efficiently while cutting costs. He uses the tool to automate a daily email that summarizes top AI usage tips from industry influencers on YouTube. After noticing the task was using up “a lot of tokens,” Neeb asked Claude if it could rework the automation to operate more cheaply. Claude revised the workflow, cutting token usage by 80%.