Across the financial sector, AI adoption has taken hold through spending on tools ranging from customer-facing chatbots to dedicated internal research units. Across enterprise environments, attention has moved away from capability questions toward the practicalities of running AI at scale — particularly what that rollout actually costs. Wherever peer institutions moved in the same direction, their own expenses climbed at a pace that proved difficult to contain.
Two pressures converge here: the practical difficulty of deploying AI at an organisational scale, and the way AI vendors have been reworking how they charge for their products. Javelin Strategy & Research, in a report on AI cost management put together by technology and infrastructure analyst Matthew Gaughan, documented how banks are reconsidering their approaches to artificial intelligence deployment amid evolving economic pressures. A narrower, more deliberate allocation of resources has taken shape in place of any sweeping rollout of the technology.
Navigating a shift in direction, the banking sector has accumulated hard-won insights from earlier rounds of technology adoption that now stand to inform what comes next. How organisations have absorbed cloud computing — and woven FinOps into their broader operations — may point toward a workable model for keeping AI consumption and its associated costs in check.
Early in their adoption of artificial intelligence, banks took an approach that mirrored much of the broader corporate landscape, directing resources toward productivity gains at the level of individual employees. On an as-needed basis, staff drew on digital tools to handle routine work like slide formatting and drafting correspondence. Wider deployment of AI tools across organisations followed as the underlying models grew in capability.
Revenue from corporate clients proved more attractive to artificial intelligence developers than what consumer chat products could generate, driving the industry's top research organisations to broaden their suite of business-facing tools. Drawn by the prospect, a wave of AI firms rolled out a staffing model built around forward-deployed engineers placed directly alongside clients. Partnering directly with client organisations, these staff members design and roll out AI solutions tailored to each group's specific needs. Within the banking sector, fraud-detection capabilities and AI-driven personalisation tools have taken shape through this kind of methodology.
Uptake across organisations has extended into areas such as agent-driven workflows, workplace productivity tooling, and software engineering environments. Pricing structures that AI providers put in place during the early period played a part in driving that growth, Gaughan noted, explaining that the technology was heavily underwritten by venture funding at the outset, which kept access costs low — typically a flat monthly fee that did not meter usage by token — and that adoption tended to begin with individual employees rather than rolling out as a formal, organisation-wide tool.
"The terminology wasn't part of how users spoke at that period," the speaker noted, adding that restrictions on how much they could use it remained a separate possibility. The arrangement persisted, he noted, though charges were structured around something other than a per-token usage model.
Pivoting toward enterprise clients brought AI providers face-to-face with ballooning computational requirements, all while the pressure to recoup the capital already poured into infrastructure buildout and model development remained a constant backdrop. Mounting competitive and financial forces nudged the industry away from flat subscription fees toward consumption-driven pricing structures, among them token-based arrangements that bill customers in proportion to the computing resources they draw upon. Pricing structures at a number of these companies blended flat-rate subscription plans with charges tied directly to actual usage levels.
Under a typical arrangement, an organisation draws on a fixed monthly allowance of credits, with any usage beyond that threshold triggering charges tied to the volume consumed. At enterprises operating above a certain scale, what a customer pays has grown increasingly dependent on how much they actually use — measured in tokens and overall consumption volume.
Those pricing and structural changes carry tangible weight for enterprise buyers, pushing many organisations to reassess where AI spending goes and how deployment frameworks get built. At one institution — whose identity has not been disclosed — AI-related expenditure reached $500 million across a single month, a figure Gaughan traced to teams that had grown accustomed to flat-rate, subsidised access and never adjusted their oversight practices once pricing shifted to a token-based structure; the resulting cost increases led a number of banks to scale back their programmes, and while that case sits well outside what is typical, the broader adoption of AI tools across the banking sector is placing growing strain on both operational capacity and financial planning, reflecting the combined weight of institution-wide deployment complexity and the overall outlay such technology demands.
Drawing on the experience of cloud computing adoption, some observers argue that organisations now revisiting these technology deployments would benefit from examining what earlier rollouts revealed — though one voice in that conversation noted the findings themselves are still taking shape. Gaughan noted that while banks have spent roughly a decade moving cloud infrastructure into wider operational use beyond data storage, the discipline that took shape during that rollout — FinOps, a term blending finance with developer operations — offers a transferable model for AI deployment, one oriented toward extracting business value from technology and embedding that priority across teams, which when applied to AI spending could give banks a structured basis for evaluating model selection, tracking token consumption, and assigning budget accountability.
Beyond simply tracking adoption rates or usage figures, the approach gives organisations a framework for weighing AI spending against the tangible returns those tools actually deliver. Despite carrying a cross-functional mandate from its inception, FinOps has tended to operate at arm's length from the wider technology organisation.
Gaughan noted that as AI extends across every function of an organisation, each business unit carries a stake in how that technology is governed — and that without active monitoring and a deliberate push toward efficiency, the associated costs can grow well beyond what any budget anticipates; FinOps, in that context, spans the full technology stack, from infrastructure at the base through to the application layer at the top. Organisations that weave FinOps disciplines into their AI spending decisions stand to navigate the tension between technological advancement and financial discipline more effectively, particularly as pricing structures across the industry continue to shift, according to the report.
Gaughan argued that where many organisational leaders have come to view artificial intelligence chiefly through a productivity and cost-cutting lens, banks and other financial institutions face a more layered imperative — one requiring AI to be addressed simultaneously as a matter of technology infrastructure and of broader business strategy, a dual framing that had little coherent operational or cultural foundation during the period when AI adoption was still taking shape. He noted that this awareness allows teams to navigate blind spots more effectively, assign the right weight to competing resources, and take clear ownership of budgets tied to various AI rollouts — and that doing so builds a layer of accountability around decisions that might otherwise have drifted without direction.
Those converging pressures, Gaughan noted, explain why FinOps has taken on such weight, and why the shifting cost structures around artificial intelligence at banking institutions have brought the industry to its current juncture.
