How CFOs can keep AI costs down

Managing the costs of AI has made CFOs’ lives a lot more complicated.

Many software providers with AI tools have abandoned seat-based pricing, a formerly fairly predictable cost for their customers. This change “makes it harder for us as CFOs because now we suddenly are dealing with consumption-based or performance-based pricing,” Matthias Steinberg, CFO of MindBridge, told CFO Brew.

Token-based pricing “makes you nervous,” Patrick Villanova, CFO of BlackLine, an accounting software provider, told us. “I don’t know what 10 tokens equals in terms of an outcome. I don’t know what the value of that is.” Now, he said, “you have to track the rate of consumption of these tokens very carefully. You have to make sure your engineers are being very cautious about how they consume Claude, how they use it.”

As CFOs at SaaS companies that serve the Big Four accounting firms and Fortune 500 companies, Steinberg and Villanova see both sides of the AI pricing equation. They shared how their organizations are keeping down AI costs.

Know outcomes before negotiating. When negotiating with software vendors, “always define the outcome first,” Villanova suggests. “People always start with price. Never start with pricing.”

Instead, he suggests defining what you want the AI or software to do: “What are you trying to achieve? What is that worth to you?” he said. “Once you have that North Star then you can justify the cost of it, and you work backward into negotiations with the vendor.” Setting a price on your desired outcome gives you a “hard line on what you’re willing to pay,” he said, and the vendor must either meet it or convince you that you’re undervaluing what they can deliver.

Help staff. At MindBridge, a financial risk intelligence platform, staff have access to a wide array of AI tools, including various models of Claude, ChatGPT, and Copilot, and “specialist modules” for finance, legal, and law, for example. The company educates employees on the costs and appropriate use of the tools at all-hands meetings, Steinberg said. “The users need to be aware that these are very expensive tools to use and that there is a usage-based pricing behind it.”

Teams should also know that they don’t need to use the most powerful model or setting for every task, Steinberg said. For instance, he said, Copilot is often cheaper than Claude but for some jobs works just as well. “For many, many applications, no one needs to be using the most cutting-edge model,” he said.

Sharing best practices for AI use also helps. “That’s the type of learning and exchange and just muscle memory the organization needs to build,” Steinberg said.

Budget for experimentation. Steinberg advocates setting aside a portion of the AI budget for learning and exploration. “Allocate a certain amount of budget that gives you enough space to experiment and for the company to learn,” he said, “based on the assumption that becoming AI native as a company and as a team is a strategic imperative.” That’s something he’s done at MindBridge and it’s making him more knowledgeable about costing, he said. The experimentation budget might not have a readily measurable ROI, “but we know this is absolutely required for us to learn how to use these tools, and that includes how to create transparency on the cost, how to learn where to actually spend the money.”

Routing controls. Soon, Steinberg believes, companies will need to implement cost control infrastructure around AI. MindBridge is developing its own digital cost controls in the form of a routing layer, “an abstraction layer between what the user wants to do and the choice of which model and which configuration is actually used,” he said. The routing layer will automatically direct users to the most cost-effective and appropriate AI models for the task they want to accomplish. “I’m deeply convinced every company soon will have” a routing layer, Steinberg said.

In the future, he added, companies might decrease their dependence on the largest AI providers by using open-source models and mini-LLMs, which are smaller and cheaper to run, but “can still give you 80%–90% of the performance” on certain tasks, he said.

Relief ahead? Both Steinberg and Villanova predict that AI costs will eventually drop. “There will be efficiencies, there will be competitive pressures, prices will come down,” Steinberg said.

“I feel fairly confident that this is not like a monopoly,” Villanova said. “Token costs will get commoditized over time. And if one vendor’s pricing goes up, there’s four or five other vendors out there that you can move to.”

The CFOs were prescient: Soon after they spoke with CFO Brew, two less-pricey AI models—Inkling by Thinking Machines and Kimi K3 by Moonshot AI—hit the market.