The conversation around Agentic AI is often dominated by speed. New capabilities emerge almost weekly, and organisations are under pressure to move quickly. Yet, Anthropic’s Dario Amodei argues in his open letter ‘We Must Pace the Frontier’ that innovation should progress at a pace where governance, safety, testing, and oversight can keep up.
For banks, the real question is not how quickly they can deploy AI, but how effectively they can scale it while maintaining observability, auditability, trust, compliance, and control. In banking, trust remains the most valuable currency.
Composable architecture as the first step
Before introducing AI, banks need foundations that are flexible, resilient, and future-ready. This starts with breaking apart monolithic core systems and embracing a composable architecture.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
A composable architecture delivers the greatest value when supported by a standard framework for integration and interoperability. This is where BIAN plays a critical role. BIAN provides a standardised interface that connects AI agents to banking systems. BIAN-based tools become portable across banks and technology platforms. This allows AI agents operating within approved workflows to interact with banking services consistently, reducing vendor lock-in and ensuring AI capabilities can scale and be reused without requiring custom integrations for every core platform. Banks that modernise their technology stacks are better positioned to integrate AI safely and incrementally. They are free to choose the technologies that best fit their needs and adapt more easily as AI capabilities mature.
Predefined workflows: efficiency without compromising control
The next stage is defining how the bank operates. Every process should be documented, governed, and assessed. Workflows orchestrate interactions across systems, humans, and AI agents. Not every process needs to be fully automated. Banks can combine automated steps with human reviews, approvals, and collaboration where required, ensuring the right balance of efficiency, oversight, and control.
Workflows establish predictable outcomes and reduce complexity by ensuring every action follows a governed path. Human decision-makers remain in, or above, the loop where required.
The goal is increased efficiency without sacrificing control.
Conversational AI
Once processes are governed, banks can introduce AI into customer and employee interactions. When a customer says, “I’ve lost my wallet,” AI can recognise the request and map it to the appropriate workflow, such as a lost or stolen card process. It can then collect information conversationally, replace complex forms with guided interactions while validating the data required to initiate the process.
Importantly, AI is not making banking decisions. It is identifying intent, gathering information, and launching predefined workflows. Conversations add value where flexibility and discovery are needed, while structured forms remain essential for precision, validation, and compliance. The result is a better customer experience, faster service, and lower operating costs.
Agentic AI with guardrails – always observable, auditable, repeatable, deterministic
The real promise of Agentic AI emerges when intelligence becomes embedded within operational workflows. At this stage, AI acts as a governed participant in a process. Banks need strong guardrails, security, predictability, and observability to prevent data leakage, manage risk, and maintain compliance.
It’s important to recognise that today’s AI agents are not inherently deterministic. They can suffer from hallucinations, and we cannot guarantee that they will produce the same results from the same inputs because they are powered by probabilistic models. Left unchecked, this can introduce operational, legal, and compliance risks. This is why banks need governed workflows around AI agents. While workflows are deterministic and follow predefined rules, AI agents provide reasoning within those boundaries. Workflows ensure that AI agent outputs are validated, monitored, and acted upon in a controlled manner, enabling banks to benefit from AI while maintaining predictability, observability, auditability, and compliance. AI may review documents, classify information, summarise cases, assess risk indicators, and provide recommendations. It can also reason over approved, read-only banking data, including customer, account, product, and case information. This richer context improves decision support without granting authority to transact.
As maturity increases, AI agents can dynamically select from pre-approved workflow activities. For example, an agent may determine whether a case requires fraud checks, additional documentation, specialist escalation, or customer confirmation. However, every action remains bounded by workflow definitions, governance policies, approval requirements, and compliance controls.
The key principle is that AI does not operate independently. An AI agent should be treated like an employee: every action must remain auditable, repeatable, and accountable.
Autonomy you can trust
Autonomy should be viewed as the destination, not the starting point. Only after extensive validation should banks permit higher levels of autonomous execution. By then, governance frameworks, workflow controls, monitoring, audit trails, and accountability mechanisms will already have been proven.
Future banking platforms may support dynamic agent ecosystems in which AI discovers capabilities, coordinates specialised agents, and orchestrates execution across systems. Yet even in these advanced environments, human oversight, validation, and end-to-end traceability remain essential.
The ultimate goal is not an AI-controlled bank. It is a bank in which AI operates safely. By pacing the frontier responsibly, banks can unlock the benefits of AI while preserving the trust that underpins the industry.
Javier Sabino, CTO, UST FinX
