AI on the Balance Sheet: How Banks Lost the Luxury of Experimentation, points report

By 2026, the gap inside banking is no longer between institutions that “believe in AI” and those that do not. Most executives will tell you they are investing. The real gap is between banks that can take a model out of a pilot environment and make it survive contact with legacy systems, regulatory workflows, and day-to-day accountability.

For much of the early 2020s, AI sat at the edges of the organisation. It was funded as an innovation programme, presented as a long-term capability, and measured through activity rather than outcomes. That arrangement lasted because the penalty for delay was low. Banks could afford to run pilots that never graduated, because competitive pressure moved slowly enough and internal inefficiency could be hidden inside broad margin pools.

That room has narrowed across most markets. Digital-first challengers expanded quickly in regions where distribution is mobile and customer acquisition is cheaper, and incumbents faced a more familiar problem in a less familiar form: speed became a competitive variable. Regulators, meanwhile, did not converge on a single global approach; they diverged, creating friction for cross-border scale while also forcing clearer governance at home. Add tighter investor tolerance for ambiguous “transformation” spend, and AI stopped being treated as a future promise. It started being treated as something that must be absorbed into operations and linked to measurable business outcomes.

This is the frame behind an executive insights report released by Dyna.Ai, developed with GXS Partners and Smartkarma, which focuses less on what AI could do in theory and more on why most banks still struggle to convert AI investment into revenue in practice. The central claim is not that banks are under-investing. It is that many institutions have built large portfolios of pilots without building the operating conditions required to scale them.

Investment is rising, but impact is uneven

The report points to the scale of projected spend as context: BFSI AI spending is expected to rise from about $35 billion in 2023 to roughly $368 billion by 2032. The number is useful, but only as a reminder that the industry has moved beyond curiosity. What matters more is how banks describe success.

The research notes that 77% of financial services executives report positive ROI within the first year of AI initiatives. That statistic sounds decisive until you ask what it means. Early ROI often appears in places where measurement is easiest and risk is lowest: small automation wins, contained pilots, and isolated improvements that do not require deep integration into core systems. These returns can be real, but they do not automatically compound, and they rarely translate into enterprise-wide financial impact unless the institution builds clear ownership from deployment through adoption and operational results.

Agentic AI exposes this gap more aggressively. Only a small minority of organisations using agentic systems report significant, measurable ROI, which suggests that the barrier is not access to models but the organisational ability to manage multi-step systems that touch multiple controls, teams, and workflows. In other words, the bottleneck moves from “can we build it?” to “can we run it safely and profitably at scale?”

Tomas Skoumal, Chairman and Co-founder of Dyna.Ai, frames the problem as one of accountability rather than experimentation: banks believe they are progressing because they have activity, but the distance between activity and measurable business outcomes is wider than most executives expect.

The revenue lever is not “AI,” it is operationalised personalisation

One of the report’s more concrete findings is that banks that successfully operationalise AI-driven personalisation can see up to a 6% revenue uplift. That figure matters because it changes where AI sits on the balance sheet. Personalisation is not a lab capability; it is a revenue behaviour. It changes cross-sell timing, offer relevance, retention, and the economics of digital service.

The report’s implication is not that personalisation is new, but that it becomes financially meaningful only when the model output is integrated into real channels with feedback loops, governance, and measurable accountability. Without that, personalisation stays stuck in “recommendations” that do not change customer outcomes or business outcomes.

This is also where the interviews in the report become more revealing than the market forecasts. William Hahn of GXS Partners notes that many executives underestimated what it would take to scale AI beyond pilots, and that this has pushed banks toward partnerships where execution and accountability can be shared rather than fragmented. The shift is not “outsourcing AI.” It is moving toward structures where providers are paid for outcomes rather than tools, because tool adoption does not guarantee impact.

Regional frontlines: the same problem, different pressures

The report focuses on Southeast Asia, Latin America, and the Middle East because these regions are where competitive pressure and infrastructure conditions force a clearer answer to the question banks often avoid: what exactly is AI doing to revenue?

Southeast Asia has two features that make AI commercially viable. The first is that customer interaction is already mobile-first, which means models can be deployed into the channels where behaviour occurs. The second is that the region carries a large MSME financing gap, estimated in the hundreds of billions of dollars, which creates a demand-side reason to improve underwriting and servicing. DBS is the visible case study: the bank reported about $565 million in AI-driven economic value in 2024 across more than 350 use cases and set an internal target of about $745 million by 2025. The number is not the point on its own. The point is that AI is being accounted for as measurable value creation across the institution, rather than treated as a set of disconnected experiments.

The Middle East is shaped by sovereign ambition and regional transaction needs. PwC has estimated that AI could add about $320 billion to the Middle East economy by 2030, with financial services positioned as a central contributor. The report highlights early impact in wealth management and cross-border payments, where AI is used to scale relationship coverage, strengthen compliance workflows, and reduce friction in regional transfers. These are practical use cases tied to speed and trust, not novelty, and they align with the region’s broader push to build AI-ready infrastructure at scale.

Latin America is framed through constraint: large-scale financial exclusion alongside persistent fraud and risk pressure. The report cites more than 200 million adults outside formal financial services in the region, and describes how AI-driven decisioning and fraud prevention are being used to extend access while maintaining risk discipline. It points to institutions such as BBVA Mexico as examples of applying AI-enabled decisioning to widen inclusion without treating risk controls as optional. In markets where fraud can destroy trust quickly, “revenue growth” is inseparable from “revenue protection,” and AI’s commercial value often shows up in reducing false declines, preventing losses, and enabling digital transaction volumes to scale.

Why pilots don’t scale

Across regions, the obstacles look familiar. Data fragmentation prevents a coherent customer view. Governance uncertainty slows deployment, especially when policy is interpreted manually rather than embedded into systems. Adoption friction appears when model outputs don’t fit frontline workflows, so usage decays even if the model is technically strong. Integration timelines remain the most expensive tax: even with a proven pilot, connecting models to core processes can take months, and those months are where market windows close.

The banks that move faster tend to embed governance early, reduce integration drag through API-first patterns, and treat feedback loops as part of production rather than a later enhancement. They also change incentives. Results-as-a-Service models, where providers are paid for measurable outcomes, are one attempt to force accountability through contract structure rather than internal promises.

Dyna.Ai positions itself in this space as a provider willing to take responsibility from domain-specific models through agentic agents and applications to operational results. Whether the market converges on that model will depend on how banks negotiate control, risk ownership, and long-term capability, but the direction is clear: activity is no longer a proxy for progress.

What changes when AI is treated as operational reality

By 2026, the question is no longer whether AI can produce value in banking. The question is whether the institution can make that value repeatable, governable, and scalable across the messy reality of banking operations.

In many banks, AI still lives in pilots because pilots are where accountability is lowest and uncertainty is most tolerated. The banks that are breaking through are doing the opposite: attaching AI to specific revenue outcomes, building ownership through to operational results, and designing execution structures that reduce the friction between “it works” and “it ships.”

That is not a story about hype. It is a story about whether banks can change their operating behaviour fast enough to match the markets they now compete in.