Summary

• Banks are moving from isolated AI pilots towards embedded operational systems across payments, fraud, compliance, customer service, and software engineering.
• The next stage of banking AI depends less on individual models than on governed data, modular architecture, orchestration, and integration with core systems.
• Human supervision, auditability, explainability, and observability will shape whether AI becomes resilient banking infrastructure or another layer of operational risk.

By Tamsin Crossland, principal AI architect at Icon Solutions

The banking industry has entered a new phase of AI adoption, as the experimental use cases and isolated pilots of the past few years begin to give way to enterprise-wide AI industrialisation. Rather than sitting at the edges of the organisation, artificial intelligence is being embedded into the operational, technical, and control layers that banks rely on to run payments, manage risk, serve customers, and build software.

According to recent research from KPMG, active use of AI across finance functions has more than doubled from 30 percent in 2024 to 75 percent in 2026. Across banking, the same pattern is becoming visible in payments, fraud prevention, compliance, customer service, software engineering, and operational support.

The conversation has therefore moved from proof of value to controlled deployment. Banks are asking how AI can be run safely, sustainably, and repeatedly across production environments, rather than whether the technology can generate useful outputs in a limited trial. For leading tier-one banks, AI is becoming less like another innovation programme and more like a foundational capability that will shape the next generation of banking infrastructure.

That shift is forcing institutions to rethink their technology architecture. As AI becomes central to operational efficiency, regulatory compliance, and competitive differentiation, success will depend on flexible, resilient infrastructure that allows new capabilities to be deployed quickly without adding another layer of avoidable complexity.

The operational AI boom

Although customer-facing AI assistants continue to dominate public attention, the deeper transformation is happening behind the scenes across banking operations. Financial institutions are embedding AI into payments, compliance, fraud operations, and software engineering not simply to test the technology, but to secure measurable gains in productivity, resilience, and process quality.

This includes the move into the next phase of AI implementation: agentic AI. Rather than deploying isolated tools, institutions are experimenting with autonomous AI agents capable of managing multi-step workflows, retrieving and synthesising information, interacting with enterprise systems, and escalating decisions where judgement or approval is required.

In practice, this points towards an operating model built around human-supervised automation, rather than full autonomy. AI agents are beginning to manage whole operational workflows, coordinating activities across systems, executing routine actions, documenting outputs, and escalating exceptions to employees who remain accountable for approvals, oversight, and regulatory judgement.

As this develops, AI is also moving beyond copilots and chat interfaces towards becoming part of the bank’s operational control infrastructure. The strongest applications are not isolated productivity tools, but systems that interpret complex inputs, reconcile data, route exceptions, document decisions, and maintain auditable trails. In that sense, AI is becoming embedded into the control layer of the institution itself.

AI as banking infrastructure

This operational shift is being matched by a more pragmatic view of AI architecture. Rather than relying on single-model or single-vendor solutions, banks are building modular architectures that combine multiple models, orchestration layers, governance controls, and enterprise data foundations.

That reflects a broader recognition that the challenge is not the model alone, but the infrastructure required to operationalise AI safely in financial services environments. Successful outcomes depend less on model selection than on access to high-quality, governed enterprise data and the ability to connect AI systems to that data in a controlled way.

As a result, capabilities such as retrieval-augmented generation, vector databases, and semantic search are becoming central components of modern banking AI stacks. These technologies allow institutions to ground AI outputs in internal policies, processes, and operational data, rather than relying on generic model behaviour.

At the same time, banks are prioritising architectural flexibility. With AI capabilities changing quickly, institutions are seeking to avoid long-term dependence on any single provider, while maintaining the ability to switch models, combine vendors, and adapt as the market develops.

The most overhyped assumption in banking is that the model itself is the differentiator. The stronger competitive advantage lies in integration, orchestration, and operational redesign.

The barriers to scale

Although banks are making rapid progress, scaling AI across the sector remains uneven because most institutions still operate across fragmented legacy systems, inconsistent data models, and limited real-time accessibility. AI performs best when data is clean, connected, governed, and context-rich, which remains a major challenge in many banking environments.

Legacy core banking and payments systems were not designed for AI-native workflows. Integrating AI into production systems, particularly where decisions are time-sensitive, mission-critical, or subject to strict regulatory obligations, can be complex, expensive, and slow.

These constraints are creating a growing divergence between banks using AI tactically for productivity gains and those redesigning operational processes around AI-enabled architectures. The first group may secure useful efficiency improvements, while the second is better positioned to change how operational work is structured, supervised, and measured.

Governance becomes operational

Governance is becoming a fundamental part of AI infrastructure. Regulators including the FCA, the Bank of England, and HM Treasury have flagged the need for institutions to prepare for new categories of AI-enabled cyber and operational risk as frontier models become more capable and more widely deployed.

At the same time, research into advanced AI systems has intensified concern that the technology can act both as a productivity accelerator and as a systemic risk factor. In banking, that makes explainability, auditability, observability, and human oversight essential design requirements rather than compliance features bolted on after deployment.

Many institutions are already moving ahead of regulation by implementing governance standards that anticipate future supervisory expectations. That includes setting clearer accountability for AI-assisted decisions, monitoring model behaviour in production, maintaining evidence trails, and designing escalation paths where human judgement remains necessary.

What AI banking looks like

Over the next three to five years, AI will become more deeply embedded into the operational fabric of banking, rather than existing as a standalone capability. Payments, compliance, fraud operations, and customer service will increasingly operate as event-driven systems in which AI supports real-time interpretation, exception handling, and decision preparation under human supervision.

As banks move from AI experimentation to industrialisation, the institutions best positioned to lead will not necessarily be those deploying the newest models first. The lasting value will come from strong data foundations, future-ready architecture, operational governance, and the ability to redesign work around systems that can be trusted in production.

About the author

Tamsin Crossland, principal AI architect at Icon Solutions, is a senior AI and financial services technology leader with deep expertise in generative AI, payments, and fraud prevention. She has led the development of a generative AI platform and co-authored a whitepaper with one of the world’s largest banks on the use of AI in payments. Tamsin has held senior roles at JP Morgan, was a systems architect at IBM, and served as engagement manager at one of the UK’s largest building societies. As a lead architect, she has delivered large-scale instant payments and digital transformation programmes for major European banks.