Setting limits on employee AI use

Finance leaders know this dilemma well: How do you control costs but encourage AI innovation? The two objectives can feel…at odds. AI “tokenmaxxing” loses its luster after companies see how expensive the bills can get.

Ryan Roccon, CFO of automation software developer Zapier, told CFO Brew about his company’s approach: Give employees room to experiment, monitor the results, and set individual limits based on that information.

Zapier has spent “a lot of our time and energy” measuring what each employee spends in AI tokens over a given period “by model, by tier, by use case” and compares that to others in their department or the organization.

However, Roccon said he doesn’t want to “handcuff” an AI power-user in product engineering if their large bill also comes with true benefits. They’ll still have a spending cap in case of a “runaway [AI] agent,” but that limit will be higher than that of an accountant who only uses AI to automate month-end close tasks.

“[When] you get deeper into what they are actually building, what value they are finding, you can start to put caps in based on team, or by individual,” Roccon said.

Learning from mistakes. Zapier encourages its employees to experiment with AI, Roccon said. Of course, this encouragement risks expensive accidents, like when a rogue agent racked up a six-figure bill in a couple of days.

Still, that costly mistake provided valuable insight. “It was a lesson we needed to learn, and it was a good lesson to learn because ultimately we saw people were using it, we saw what they were using it for, we saw where the edge cases were, [and] we saw where there could be abuse,” he said.

Central command. Zapier has an “AI transformation office” to help guide all AI-related decision-making. The group includes Zapier’s chief people officer, who’s also the AI transformation officer and is “responsible for all outcomes associated with AI transformation across the org,” Roccon said. It also includes folks from various functions including finance, communications, and procurement.

And if an organization doesn’t have its own version of an AI transformation office already, “I would recommend that they establish that,” Roccon advised. “You need a central center of excellence that can help define, measure, and account for those things.”

Work in progress. Determining the benefit of each AI use case is a process that requires “both quantitative and qualitative” approaches,” Roccon said. “And unfortunately, it’s very rarely going to be both, and you’re going to have to decide when one is sufficient.”

In some instances, it’s easy to measure results quantitatively, he said, as in the case of a customer support team that “effectively got rebuilt around AI.” Roccon can take support metrics like resolutions per hour and quality scores and compare those scores between AI users and non-AI users.

Other instances require more subjectivity. More engineering team output doesn’t always translate into better results, he said. Also, a quick product patch isn’t directly comparable to a “really deep new product enhancement.”

His approach on more difficult-to-measure use cases is to talk with the heaviest AI users and ask them to show off the AI tools they’ve built. In instances of suboptimal use of resources, like a fancy graphic illustration, they may ask the worker to use a less expensive AI model for that work. In other cases, like an engineer who made a group of AI agents that drastically increase coding output, Zapier may want to hold it up as a shining example for others to follow.

“You might look at it and be like, ‘Did the $5,000 they spent last week on all this code generation produce $5,000 worth of net new functionality driving customer satisfaction, engagement, adoption, etc.?’ Good luck being able to answer that directly,” Roccon said. “We’re not there, but I sure wouldn’t want to stop him because I couldn’t prove that yet.”