Employee AI use is growing; CFOs want financial guardrails

In a post-tokenmaxxing world, finance leaders are wondering how to set limits for AI use among teams and individuals in their companies. And, more simply, where to start?

Establishing limits for internal AI use is not a novel concept. As CFO Brew wrote in July, it’s a dilemma finance leaders know well: How do you control costs while encouraging AI innovation?

Field CTO Amit Kinha of DoiT, a cloud intelligence and finops platform, told CFO Brew that the “slush fund” for operational efficiency-focused AI token spend that existed last year is now a thing of the past. “This year, due to market conditions or the maturity of things like agents, it’s not enough to just say, ‘Yeah, everyone gets carte blanche access,’” he said.

Two buckets. Chris Andersen, CFO of Flexera, said he thinks about AI usage as two different buckets: enterprise productivity-focused AI and consumption-driven AI.

“We want everybody to have [AI], we want everybody to get used to it, it should make them more productive. That’s a cost that we invested in the business a while ago, and at the corporate level, we knew how much it was, we made room for it, we did that in the hopes that it would drive enterprise-wide productivity,” Andersen said.

The second bucket, consumption-driven AI, is “about allowances and token spend. And for us, at least in my experience at Flexera, the best example of that, really, is our product and engineering teams,” Anderson said. “The developers are using coding assistants, they’re using the higher capability models, and those are the tools that really generate the credit or the token consumption.”

Setting limits. Measuring the ROI of AI tools and their application is something that CFOs have struggled with for some time, but it is essential to figuring out who gets what AI allowance in a company, Kinha said.

“The challenge I think for a CFO specifically is that, they’re getting the bill part of it, and it’s really hard to quantify the gain part of it, like the value part of it,” Kinha said. “The conversation is quickly turning to, forget the agents helping chatbots for customer support, the individual users and our company employees, how do we think about what the ‘limit’ should be, if any?”

At Flexera, at the macro level, “the business owns the use case, finance sets the financial guardrails, and the IT team makes certain that that consumption is visible, it’s safe, it’s controlled,” Anderson said.

At the company’s operational level, setting limits “starts with some form of hypothesis around, ‘What’s the expected user population? How heavy are they going to be using the tool? What’s the volume, what’s the model or service?’” Anderson said.

Funnel AI. Appearing on a panel on Oct. 22 at Campfire.ai’s Basecamp summit in San Francisco, Steve Sidhu, a controller at Clay, a go-to-market platform, advised finance leaders to start broad with their AI allowances, then narrow them down over time.

“With these platforms, we want to set guardrails in place, and then slowly over time bring guardrails in closer and closer, while allowing experimentation,” Sidhu said on a panel about finance tech stacks. In Clay’s experimentation, in some areas, “we’ve probably gone too far and had to pull back,” Sidhu said.

But he cautioned against letting missteps keep people from tinkering. Organizations need to strive for a balance, he stressed

“Maybe two years ago, thinking about connecting a cloud to your ERP would be a little scary for most people,” Sidhu said. “And I would say most people sitting in this room probably have Claude in their ERP now, and are doing a lot of their financial work in Claude or OpenAI and using those tools.”