2026 banking and capital markets outlook

Five steps banks should consider to move beyond isolated AI projects

The year 2026 could be pivotal for banks as they aspire to become fully AI-powered. Currently, AI implementation within banks is often throttled by brittle and fragmented data foundations, mounting compliance demands, outdated legacy systems, and internal resistance to change. Many AI initiatives are stuck in isolated proofs of concept, marked by weak governance, duplication, and uneven impact.

Many bank executives also seem to be grappling with unrealistic productivity expectations and facing increasing pressure to demonstrate tangible results. Despite large and growing AI budgets over the past two years, most US banks have only achieved sporadic tactical wins rather than true strategic transformation.⁴⁵ Our review of the top 40 US banks reveals predominantly “reactive,” siloed efforts that yield inconsistent value.⁴⁶

Reframe a clearer and more unified AI vision and strategy

Until now, most banks have generally taken a federated and patchy approach to AI, especially generative AI. While many have experimented, adoption often lacked an overarching vision. Is the primary aim to drive efficiency, accelerate innovation, or strengthen risk management and resilience? Without a unified vision, banks may struggle to identify scalable AI opportunities and measure progress against key performance goals.

To date, only a handful of institutions have articulated a cohesive, firmwide AI strategy where every piece fits together and operates in unison. To succeed, the vision should spell out concrete outcomes; recognize risks, costs, and human implications; align with the bank’s broader mission; be communicated consistently across all stakeholder groups; and be underpinned by disciplined funding.⁴⁷ Done well, this could prevent the sprawl of disconnected pilots and channel resources toward initiatives with the greatest strategic impact.

Establish clearer ownership and governance for AI

Banks should have clear ownership across the AI life cycle, yet accountability is often fragmented or absent.⁴⁸ Approaches also vary in how employees can access and use AI tools, making it important to define which responsibilities sit with a central team and which reside within business units.

For most banks, a hub-and-spoke model could be an optimal choice. This model can help ensure that the needs of different business lines are adequately managed, anchored by a central unit like an AI center of excellence.

This central entity can help drive quality across the enterprise and uphold AI governance standards while serving as the operational hub for AI adoption—maintaining a living roadmap for execution across the enterprise. In addition to developing the AI strategy, it could also be responsible for reference architecture, standards, shared assets, and MLOps or LLMOps⁴⁹ services to help ensure interoperability. Beyond governance, the center of excellence could focus on training, playbooks, and knowledge sharing, and help support delivery by operating core AI platforms.

Reassess the “build vs. buy” calculus

The build vs. buy choice is another recurring dilemma but takes on a different flavor with AI.  Many banks have adopted a hybrid model for traditional AI, like machine learning—building proprietary models while buying point solutions and platforms for less differentiated needs.⁵⁰ For gen AI, some banks have shifted the focus toward an assembly approach in which they buy the foundation model layer but build custom proprietary layers around it with data connectors, guardrails, and third-party solutions.

Beyond leveraging third-party expertise, this approach can help reduce time to market and experimentation costs. The “buy” option can also shift the risk of potential cost increases to third parties.⁵¹ Smaller banks, in particular, often have no choice but to adopt a hybrid approach because of tighter budgets, scarce talent, and lower risk tolerance.⁵²

However, the assembly approach is not without challenges. For instance, proprietary layers should be well integrated with the foundational model(s). Also, if every bank uses the same models—or the models from third parties happen to be similar—the only differentiation lies in the proprietary, bank-specific layers.

To help build their competitive differentiation with gen AI, banks should lean heavily on proprietary data. They should also be creative in where and how these models are applied: narrow, high-impact workflows could outperform sprawling moonshots. Finally, banks should invest in specialized talent like prompt or retrieval-augmented generation (RAG) engineers, evaluators, and designers who can turn models into robust systems. This is where true differentiation might lie.

Measure and track ROI with discipline

As AI scales, measuring impact can become critical, yet some senior executives find it hard to assess value beyond anecdotal or subjective metrics like hours saved or calls shortened.⁵³ Software developer productivity is perhaps one area where ROI measurement is most advanced.⁵⁴

Without standard baselines, counterfactuals, or consistent key performance indicators, benefits often rest on user claims rather than measurable financial outcomes. This can create a credibility gap, making it hard to link soft benefits to tangible cost savings or revenue gains. Many gains can also be second order: for example, shorter customer service calls may improve customer satisfaction, helping drive cross-sales—yet these effects remain hard to quantify. Gen AI can complicate the issue with claims of productivity not connected to actual costs.⁵⁵ Only 4 out of 50 banks analyzed by Evident in 2025 reported realized ROI from AI use cases.⁵⁶