Banks Race into AI as Cost and Governance Risks Mount

Banks have shed their reputation as technology laggards. But even as AI deployments multiply and early results look promising, findings from a recent survey show many are dealing with runaway token costs, the spread of unmanaged tools and IT infrastructure gaps. These threaten to impede progress. 

Royal Bank of Canada (RBC) has developed more than 200 AI models and seen its use of large language model tokens has surged more than 500% since last year. JPMorgan Chase has earmarked nearly $20 billion for technology this year, including generative AI projects, with nearly 1,000 AI use cases across the company. Bank of America, which allocates more than $4.1 billion annually to new technology initiatives, has more than 300 AI projects in development.

Long considered one of the more conservative technology adopters, the financial services industry is now among the most aggressive, with 81% of firms adopting AI in some way and 40% are in advanced stages, a University of Cambridge Judge Business School global survey found.

Projects are proceeding quickly and mostly according to plan. But findings from Nutanix’s 2026 Financial Sector Enterprise Cloud Index (ECI) tell a deeper story. The survey of IT professions found that outdated processes, organizational barriers, and infrastructure gaps are slowing the industry’s move from AI pilots to production, even as the technology itself performs.

“Banking is a good business right now, but it’s increasingly costly, and there are big rocks that have to be moved to successfully deploy AI,” said Sean O’Dowd, head of global financial services at Nutanix. 

“The rocks are what you’d expect: organizational drag and outdated infrastructure.”

Overcoming Organizational Drag

AI adoption inside financial institutions has become, in a way, a victim of its own success. Employees want tools that can speed routine work, shorten project timelines, and improve customer service. Many institutions still lack the policies and approved alternatives needed to keep that use safe. Few have rolled out tools that match the ease and capability of services employees can access directly from OpenAI, Anthropic and other GenAI providers.

That gap between what financial institutions can provide and what employees expect led 38% of ECI participants to cite process friction as a barrier to AI adoption, 34% to call organizational factors an issue and 28% to see technical limitations as a key hurdle.

Cost control is becoming equally pressing as employees use AI more often and reach for models that may cost more than the task requires.

During JPMorgan's second-quarter earnings call, CFO Jeremy Barnum called for more discipline in model selection. 

“The idea is (to) use the right model for the right purpose, be smart about open source where appropriate, and ensure that you’re getting value out of it ultimately,” he said.

A major source of variable expense is usage-based pricing, generally calculated by tokens, the small units of text AI models process in prompts and responses. Agentic systems can consume far more tokens than chatbots because they read files, call models repeatedly, and work through multistep tasks. A July 2026 J.P. Morgan Private Bank analysis found effective token prices had risen more than 60% since December.

Such increases are leading many executives to weigh token costs alongside software licenses and headcount in budget discussions, a practice sometimes called tokenomics.”

“When you have a tool that’s generating $100,000 in savings, but you spend a million on tokens to get there, what’s the true value of that?” O’Dowd said. “There’s just a massive lack of transparency into the bill that they are paying on a monthly basis.”

Rethinking Infrastructure and Controls

Other costs are harder to see.

As bring-your-own-device policies once forced banks to manage personal devices on corporate networks, shadow AI is forcing them to confront tools and projects operating outside IT oversight.

The ECI found 86% of financial services IT executives believe AI tools and agents used outside official oversight create business risk. Sixty-six percent regularly discover applications or agents implemented by non-IT employees without IT oversight.

That blind spot can create regulatory exposure. Banks may not know what customer data employees entered into outside tools, where they were stored, or whether its use complied with privacy, retention, and governance rules. They may also struggle to document controls, explain AI-assisted decisions, or tell regulators how sensitive information was handled.

Data sovereignty adds another constraint. The ECI found 79% of respondents view it as a must-have, yet 62% currently run containerized applications in the public cloud, a gap the report calls “sovereignty debt.” Banks must be able to show where data resides, who can access it, and which jurisdiction governs it. Those requirements could push some AI workloads toward private clouds or on-premises systems.

“Regulatory scrutiny has ratcheted up substantially over the last five years,” O’Dowd told The Forecast. "I can't think of such far reaching, higher level of fine and compliance hurdles, as well as mismatched requirements for mandates globally and domestically, imposed on any other sector. It's hard for me to really think of another industry where it's such a high cost of doing business."

O’Dowd said regulators are also scrutinizing the risk of banks concentrating on a small group of technology providers. It raises many concerns, especially where an outage could disrupt payments or trading.

Those concerns are reshaping infrastructure plans. In the ECI survey, 46% cited performance and 43% named data security as primary reasons for adopting containers. Regulatory compliance and risk management also influenced deployment decisions.

Building for scale

Few financial institutions have settled on a repeatable way to deploy enterprise AI efficiently, securely and at scale.

O’Dowd said leading banks are building centralized AI platforms that apply consistent governance, security and cost controls across projects. He calls it an “assembly-line approach” that can move applications from pilot to production more quickly.

"Centralized AI platforms are allowing them to take this assembly line approach and roll out, not just roll that out faster and wider, but requiring the needed governance and management on that," he said. "I've heard some of the most major banks just talk about credible enterprise AI platforms are what really separates serious players from the rest."

David Rand is a business and technology reporter whose work has appeared in major publications around the world. He specializes in spotting and digging into what’s coming next–and helping executives in organizations of all sizes know what to do about it.

© 2026 Nutanix, Inc. All rights reserved. For additional information and important legal disclaimers, please go here.