The finance sector’s relationship with artificial intelligence has moved past the point where adoption itself is the story. Investment is widespread, leadership buy-in is nearly universal, and the integration work is actively underway across the majority of organizations. The more important question now is whether that investment is producing the financial value it is supposed to produce, and the honest answer is that for most organizations it is not, at least not yet. Boston Consulting Group’s 2025 research across 1,250 global companies found that only 5% are achieving returns that justify the investment at scale. The gap between the near-universal belief that AI is essential and the narrow slice of organizations actually realizing meaningful returns is not a technology problem. It is an implementation problem, and understanding what separates the 5% from the rest is the more useful frame for finance leaders deciding how to approach the next phase of their AI strategy.
The Finance Sector’s Commitment to AI Is Not in Question
The scale of finance sector investment in AI is significant enough to serve as context for any conversation about returns. Research from UK finance leadership finds that 99% view AI as essential to their operations, and 85% of finance teams are actively integrating it rather than waiting. Investment is projected to continue rising through 2026. These figures describe an industry that has made a collective judgment that AI capability is not optional, and that judgment is difficult to dispute given what the technology demonstrably enables.
AI’s potential applications in finance are substantial and span functions that have historically been labor-intensive, error-prone, or limited in scale by the capacity of human analysts. Automation of repetitive processing tasks reduces cost and error rates simultaneously. Fraud detection systems operating on AI models can identify patterns across transaction volumes that no human team could monitor continuously, and they improve as they process more data. Predictive analytics applied to customer data, market signals, and operational metrics generates insights at a depth and speed that changes what is actionable rather than simply what is known. Cybersecurity applications using AI can detect and respond to threats with a speed and pattern-recognition capability that manual monitoring cannot replicate.
The case for AI investment in finance is well supported. The returns that the majority of organizations are failing to realize are not evidence against that case. They are evidence that realizing those returns requires more than purchasing and deploying the technology.
Why Most AI Investment Is Not Producing the Returns It Should
The BCG finding that only 5% of organizations are seeing strong returns on AI investment is striking precisely because it sits alongside near-universal adoption. The explanation is not that AI does not work. It is that the conditions required for AI to produce value at scale are more demanding than the technology purchase itself suggests, and most organizations have not yet met those conditions.
The most common failure mode is deploying AI within existing workflows rather than using AI as a reason to reconsider whether existing workflows are the right structure. An AI tool that automates a step in a process that was designed for manual execution will produce incremental efficiency gains. An AI tool deployed in a workflow that has been redesigned around what AI can do will produce results of a different order. The distinction matters because the second approach requires accepting that processes which have worked reliably for years may not be the right foundation for an AI-enabled operation, and that acceptance is genuinely difficult for organizations where those processes represent accumulated institutional knowledge and established ways of working.
The human capability gap is the second dimension where most implementations fall short. AI tools require people who understand what they are capable of, how to interpret their outputs, and when to trust or question what they produce. Finance functions in particular require judgment about when AI-generated analysis is reliable enough to act on and when the underlying data, model assumptions, or edge cases in a specific situation require human scrutiny. Building that judgment takes time and deliberate investment in development, and organizations that deploy AI without investing in the people operating it tend to find that the tools underperform relative to their potential because the human layer needed to get full value from them is not present.
What the Organizations Achieving Returns Are Doing Differently
The organizations in the 5% share a set of replicable characteristics, though not without effort, by organizations currently outside that group.
The first is treating AI as a strategic foundation rather than a functional add-on. The organizations realizing strong returns have asked the harder questions: how should work be allocated between human employees and AI systems, where does human judgment add value that AI cannot replicate, and how should resources be redirected to the activities where human contribution is highest. Those questions produce different answers than the simpler question of which existing tasks can be automated, and they produce different results. Starting with specific, well-defined integration points where AI fits naturally into existing operations and expanding from that foundation is more reliable than attempting broad transformation before the organizational capability to manage it exists.
Workflow reinvention rather than workflow automation is the second distinguishing characteristic. The organizations achieving strong returns have not simply applied AI to their existing processes. They have used AI deployment as an occasion to examine those processes from the ground up and redesign them around what is now possible. AI inventory monitoring that operates continuously without human intervention is not valuable primarily because it replaces a human task. It is valuable because continuous monitoring enables responses to conditions that periodic human review would have missed. The value comes from what the redesigned process makes possible, not just from the labor cost of the task that was automated.
Talent development treated as a parallel investment to technology investment is the third characteristic. AI tools produce results proportional to the capability of the people operating and interpreting them. Finance functions where employees understand how to work with AI outputs, where to apply skepticism, and how to use AI-generated analysis as a starting point for more sophisticated judgment rather than a final answer, consistently outperform functions where the same tools are operated by people without that development. The specific development priorities that produce this capability include building genuine AI literacy rather than surface familiarity, reinforcing the analytical and judgment skills that AI does not replicate, and addressing the legitimate concerns about role change that employees experience during significant technology transitions. Organizations that manage the human dimension of AI deployment carefully retain the people with institutional knowledge needed to operate AI systems effectively.
Infrastructure Investment Determines Whether AI Can Scale
The returns that are visible in the early stages of AI deployment are often produced by tools that operate on existing data infrastructure. Scaling those returns requires infrastructure that was designed to support AI workloads rather than simply tolerate them.
Data quality, accessibility, and governance are the foundation on which AI systems operate, and the organizations that have invested in getting those foundations right before attempting to scale AI deployment are consistently better positioned than those that have tried to solve infrastructure problems while simultaneously expanding AI use. AI systems that operate on incomplete, inconsistent, or poorly governed data produce outputs that require more human review to be trustworthy, which reduces the efficiency gains the deployment was intended to achieve.
Infrastructure providers have developed pre-built and customizable options specifically designed to support AI deployment at scale, with integration, scalability, and security built into the architecture rather than retrofitted after the fact. For finance organizations where data security and regulatory compliance are non-negotiable, the infrastructure layer is not a component that can be treated as a cost to minimize. It is the prerequisite that determines whether AI investment can produce returns at the scale that justifies the broader commitment.
The finance sector’s AI transformation is real, the investment is substantial, and the potential returns are significant enough to justify continued commitment. The organizations that will capture those returns are the ones that approach implementation with the same rigor they would apply to any major capital allocation: clear objectives, honest assessment of current capability gaps, parallel investment in the human and technical infrastructure needed to support the strategy, and the willingness to redesign how work is done rather than simply automate how it is currently done.
