Banks have moved quickly to embrace artificial intelligence, investing in everything from sophisticated chatbots to in-house research teams. But as AI spreads across the enterprise, the challenge is shifting from what the technology can do to how much it costs to deploy at scale.
As other financial institutions have followed suit, many have encountered rapidly rising costs. This reflects not only the challenges of scaling AI across the enterprise, but also the evolving pricing models of AI providers.
As Matthew Gaughan, Tech & Infrastructure Analyst at Javelin Strategy & Research, detailed in the As AI Costs Grow, Here’s How Banks Can Manage the Economics report, these changes have forced banks to reconsider their approach to AI implementation. Rather than deploying the technology broadly, they are looking for ways to invest more strategically.
As they chart a new course, banks can draw lessons learned from previous technology implementations. Namely, the adoption of cloud computing—and the enterprise-wide incorporation of FinOps—offers a potential blueprint for managing growing AI usage and expenses.
From Individual Productivity to Enterprise Adoption
Initially, AI adoption at banks followed a trajectory similar to that of many other companies, with a narrower focus on individual productivity gains. Employees used technology on an ad hoc basis to simplify tasks such as formatting PowerPoint presentations and composing emails.
As leading AI models became more capable, adoption become more formalized across organizations. For AI providers, enterprise customers offered a more lucrative market than consumer-facing chat interfaces, prompting leading labs to rapidly expand their business offerings.
To capitalize on this opportunity, many AI companies also introduced a new model for customer engagement: forward-deployed engineers. These employees work directly with organizations to develop and implement customized AI solutions.
In banking, this approach has helped produce targeted fraud mitigation tools and personalized AI agents. It has also supported broader adoption across institutions, spanning agentic workflows, productivity tools, and software development.
This expansion was facilitated, in part, by the pricing models AI providers initially offered.
“On the economic side, the service itself was in large part subsidized from early funding, and enabled users to access it,” Gaughan said. “In the early stages, this was mostly on a personal level—not necessarily professionalized throughout the organization—at a low-cost, flat subscription that didn’t necessarily involve tokens. That wasn’t in the vernacular of users at that time.”
“That’s not to say that they wouldn’t be limited on their usage. That still existed, but in terms of payment, they weren’t directly paying on a per-token basis,” he said.
The Pricing Evolution
As AI providers shifted toward enterprise customers, they faced growing computing demands while continuing to seek returns on substantial investments in infrastructure and model development.
These pressures contributed to a shift away from purely subscription-based services toward usage-based pricing, including token-based models that charge customers according to how much computing capacity they consume.
Some providers adopted hybrid models that combine subscriptions with usage-based fees. For example, a business might receive a set number of credits per month before incurring additional charges based on consumption. For many larger organizations, pricing has become tied to usage and token consumption.
These changes have significant ramifications for enterprise customers, forcing organizations to reevaluate both their AI spending and deployment strategies.
At the extreme end of the spectrum, an unnamed company reportedly spent $500 million in a single month accidentally, after forgetting to adjust AI usage limits for employees in light of the pricing evolution.
“There was probably a disconnect, people who had been using the technology early on where it was effectively subsidized and there wasn’t this token-based model got used to that and were not paying attention to usage,” Gaughan said. “But after these cost increases, focus on that became more important; you saw a lot of banks pulling back.”
Lessons from the Cloud
While that incident is an outlier, the AI build-out and adoption at many banks has become a significant strain, due to both the complexity of deploying AI across the institution and the overall expense.
As institutions reassess these implementations, there are lessons from previous technology integrations that they should consider, most notably from the recent rollout of cloud computing.
“These are still new lessons. Cloud adoption in earnest at the banks and use in more functions beyond just data storage is a little over a decade old,” Gaughan said. “But there are some relevant lessons that could be learned from those previous cycles and applied to the AI build-out.”
“In the case of the cloud build-out, something emerged at companies called FinOps, which is a portmanteau of finance and developer operations,” he said. “FinOps is focused on maximizing the business value of technology and spreading that ethos throughout the culture of the business, specifically the technology team.”
Applied to AI, FinOps can give banks a more comprehensive framework for making decisions about model selection, token consumption, and budget accountability. It can also help institutions evaluate AI investments against their business value rather than focusing solely on adoption or usage.
Although FinOps has had a cross-functional remit from the start, it has still often been separated from the broader technology organization. This approach should change given that AI now touches every part of the business, every business unit should be involved.
“That scale and breadth also means the costs are going to be high and can easily get out of control if they’re not being monitored, and if they’re not finding ways to make a more efficient use of technology to maximize its value,” Gaughan said.
“FinOps operates across the entire stack from infrastructure all the way up to application,” he said. “It’s less about cutting costs and more about finding ways to allocate resources properly in a way that saves on cost but doesn’t sacrifice the benefits that AI could provide.”
The Changing Economics of AI
Financial institutions that embed FinOps into AI decision-making will be better positioned to balance innovation with cost control as the technology and its pricing models continue to evolve.
All in all, the implementation of FinOps exemplifies how financial institutions should continue to approach AI: through a broad lens. Due to its capabilities, many leaders have begun to view AI solely as a business function built to boost productivity and cut costs.
“Especially at banks, it’s important that AI is considered as a tech and infrastructure consideration as much as it is a business one,” Gaughan said. “In the early stages of AI adoption, there weren’t any cohesive operational frameworks or cultural practices to instill this ethos of maximizing the business value of technology.”
“That’s where that all comes together. And having that in mind helps them see around blind spots, know how to prioritize different resources, and establish ownership over budget for different implementations of AI,” he said. “It creates a level of accountability on some of these things that might have previously been left to fester on its own. That all feeds into the importance of FinOps, and how the changing economics of AI at banks has gotten us to this point.”
