Across the financial sector, AI tools — from customer-facing chatbots to dedicated internal research units — have drawn substantial investment in a relatively short period. Wider adoption of AI across corporate operations has moved the central conversation away from technical capability and toward the practical economics of running these systems at scale.

A Javelin Strategy & Research report on AI cost management — authored by Tech & Infrastructure Analyst Matthew Gaughan — found that shifting economic pressures are prompting banks to reassess how they deploy artificial intelligence. The focus has shifted toward targeted investment rather than rolling out the technology across the board.

Two pressures are converging: the operational difficulty of deploying AI at scale across large organisations, and the way AI vendors have been adjusting how they charge for their products. In each case where comparable institutions followed a similar course, the actual costs ran well above what planners had projected.

Experience from earlier rounds of technology adoption gives banks a practical foundation as they navigate fresh strategic directions. How companies have handled the shift to cloud computing — and woven FinOps into their operations at scale — may offer a workable model for keeping AI adoption and its associated costs under control.

Banks, in the early phases of technology uptake, tended to concentrate on boosting output for individual employees rather than pursuing broader operational change. On an unplanned, as-needed basis, staff reached for digital tools to handle routine work such as slide formatting and drafting correspondence. Structured adoption of AI tools took hold across organisations as the underlying models grew in capability.

Building out enterprise-grade tools and services drew far greater revenue potential for AI developers than consumer chat products did, pushing the top research organisations to broaden their commercial portfolios at pace. Seizing on that opening, a number of AI firms rolled out a fresh approach to working with clients: engineers placed directly on-site with customers. Tailored AI solutions take shape through direct collaboration between those staff members and the client organisations they serve.

Within the banking sector, fraud-detection utilities built around specific risk profiles and AI-driven personal finance assistants have both taken shape through this methodology. Uptake across organisations has grown to encompass agentic workflows, productivity tooling, and software development environments.

Pricing structures offered by AI providers played a role in driving that growth, according to Gaughan, who noted that venture funding had kept costs down for users in the early period, when adoption tended to happen at the individual level rather than across whole organisations, through flat-rate subscriptions that were not metered by usage volume. "That phrase wasn't part of how users spoke at the time," the speaker noted, adding that restrictions on how much they could use it remained a possibility. The arrangement persisted, he noted, though charges were not structured around individual token consumption.

As AI providers shifted their focus toward enterprise customers, the computational demands on their systems grew considerably — and so did the need to recover the costs already sunk into infrastructure and model development. A broader industry pivot away from flat subscription fees took hold as providers moved toward consumption-driven arrangements — among them token-based structures that bill customers in proportion to the computing resources they draw on.

Across a portion of the market, blended pricing structures took hold — pairing flat-rate subscriptions with charges tied directly to actual usage. Under a typical usage arrangement, an organisation draws from a fixed monthly allocation of credits, with any consumption beyond that threshold triggering extra fees. At enterprises operating at scale, what customers actually pay has shifted toward metered models driven by usage volume and token consumption.

Gaughan pointed to one unidentified institution that accumulated roughly $500 million in AI-related expenditure across a single month after internal consumption limits were left unchanged while vendor pricing moved away from earlier subsidised structures — a situation he traced to teams that had grown comfortable with low-cost access and were not monitoring usage closely enough, a discovery that led numerous banks to scale back once those higher costs took hold. While that episode stands apart from the norm, the broader push to embed AI throughout the banking sector has been placing growing strain on budgets and operations, shaped by the dual pressures of deploying the technology at institution-wide scale and managing the overall expense that comes with it. Enterprise customers now face a recalibration of both their AI budgets and how they put those tools to work, driven by the shifts these updates introduce.

Watchers of the industry draw on the cloud computing transition as a reference point, noting that how these technologies ultimately take hold across the sector remains an open question — a point underscored by at least one practitioner, who described current knowledge in the field as still far from complete.

Gaughan pointed to roughly a decade of bank experience deploying cloud technology across wider operational areas as a source of portable lessons for AI infrastructure planning — with FinOps standing out as one of the more telling products of that period. FinOps is a practice that merges financial oversight with developer operations and orients itself around extracting measurable returns from technology spending while embedding that mindset into organisational culture, especially among technology teams. When adapted to AI contexts, it offers banks a structured way to approach decisions around model selection, token usage, and budget accountability.

Beyond simply tracking adoption rates or usage figures, the approach opens a path for organisations to weigh 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 because AI now runs through every layer of an organisation — from infrastructure up through individual applications — all business units need a seat at the table, and without active monitoring and a drive toward more efficient technology use, the associated costs can grow well beyond what any budget anticipates. FinOps, he added, addresses that challenge across the full technology stack.

Gaughan noted that while many organisational leaders have come to treat AI primarily as a productivity and cost-reduction tool, financial institutions in particular need to approach it as both a technology and infrastructure matter — and that the absence of unified operational frameworks or shared cultural norms in the early phases of adoption left that broader understanding underdeveloped. He noted that carrying that awareness allows teams to identify blind spots more readily, set priorities across different resource types, and take clear ownership of AI-related budget decisions — while also building in a layer of accountability for matters that might otherwise have gone unaddressed.

Organisations that weave FinOps disciplines into their AI spending decisions stand to achieve a workable equilibrium between pursuing technological advancement and keeping expenditure in check — a balance that may prove harder to maintain as pricing structures around the technology continue to shift. Those converging pressures explain why FinOps has drawn attention, and how a shift in the financial economics of artificial intelligence at banking institutions brought the industry to its current position.