For several years, banks have been asking a relatively straightforward question about AI: Where can we use it?
That question produced hundreds of use cases from summarizing documents and supporting research to improving customer service and automating routine work.
The next question is far more consequential: What happens when AI stops assisting individual tasks and begins participating in how a bank operates?
As AI becomes cheaper, more powerful and widely available, intelligence itself is becoming less of a competitive advantage. For banks, the next challenge is coordinating AI across people, systems and workflows at enterprise scale.
Consider a disputed transaction. Today, resolving it may involve customer service, transaction records, fraud systems, identity controls, case management, payment networks and compliance procedures. An AI assistant might help an employee summarize the case.
An agentic workflow could gather records, reconcile information across systems, identify anomalies, initiate prescribed checks, route exceptions and prepare recommendations for human review.
The first approach improves a task. The second begins to redesign the workflow.
When AI begins operating across workflow boundaries, institutions must decide which work remains human, which can be delegated to machines and where human judgment should intervene. That is no longer a technology challenge. It is an operating model challenge.
As banks move from experimentation to enterprise adoption, the three factors that will determine whether agentic AI can scale are context, control and coordination: Whether agents have access to the right information, whether their authority is appropriately governed and whether people, systems and agents can operate together as one coherent enterprise. Among these, coordination may become the defining challenge.
Coordination is the real scaling challenge
AI debt emerges when organizations deploy agents faster than they develop the structures needed to coordinate them. Individual initiatives may work well on their own. Collectively, they can create a difficult environment to govern, monitor and scale.
While technology debt may be confined to the IT department, AI debt can cause reputational and business risk.
A bank with 50 AI applications can often manage them as individual systems. A bank operating thousands of agents across interconnected workflows faces a different problem entirely. It needs to understand what happens when agents interact with one another, with systems and with human decision-makers.
This is where orchestration becomes an enterprise capability rather than a technical feature.
The goal shifts from automation to coordination, making interactions across people, processes, systems and agents visible, governable and economically sustainable.
Data becomes context — and context determines judgment
Orchestration alone is not enough. Agents need context to operate effectively.
Banks do not lack data. They hold extraordinary volumes of customer, transactional, market, risk and operational information. The challenge is that much of it resides across different systems, formats, taxonomies and business units.
For many U.S. institutions, that fragmentation reflects decades of mergers, platform decisions and technology investment. Those environments were designed primarily for applications and human users — not autonomous software attempting to reason across the enterprise.
Having data is not the same as having usable context.
An agent investigating a customer issue, for example, may require information from a CRM platform, transaction history, fraud controls, product systems and previous service interactions. Each source may be accurate on its own while still failing to present a consistent picture of the customer.
Humans compensate for these gaps. They call colleagues, interpret exceptions, recognize missing information and understand institutional history.
Machines are less forgiving.
As agents move from retrieving information to recommending or executing actions, data quality ceases to be primarily an analytics problem. It becomes an operational risk issue.
The implication is that AI strategy and data modernization can no longer be treated as separate agendas.
Institutions that postpone difficult work around data lineage, identity management, metadata, access controls and data quality may find that they have sophisticated AI capabilities built on foundations that cannot support trusted decision-making.
The institutions that create advantage in the next decade will not simply deploy intelligence at scale. They will coordinate it.
Governance must move from focusing on models to decisions
Banks already operate within mature frameworks for cybersecurity, model risk, privacy, operational resilience, third-party risk and regulatory compliance. Agentic AI does not eliminate those disciplines. It changes how they intersect.
Traditional governance frameworks have focused on models and systems. Agentic environments require institutions to govern decisions.
Responsibility may be distributed across models, agents, enterprise applications, external services and human supervisors. The critical question becomes: Who is accountable for the outcome?
That question grows more important as AI moves from recommendation to execution.
An AI system that drafts a response presents one level of risk. An agent that can modify records, trigger workflows or initiate actions presents another.
Regulators are already grappling with these questions. The Financial Stability Board has identified areas of potential AI-related vulnerabilities, while U.S. banking regulators continue to examine how existing model risk and third-party risk approaches should evolve as AI capabilities advance.
The regulatory direction is becoming clearer, even as some questions remain unresolved: Greater autonomy will require greater oversight.
The next competitive advantage
Perhaps the most overlooked implication of agentic AI is economic.
For decades, software has improved efficiency by reducing marginal labor costs. Agentic systems introduce a different dynamic.
Reasoning itself is becoming a metered operating expense.
Every inference, action and escalation carries a measurable cost. When thousands of agents continuously start models, access data, and interact across workflows, inference and infrastructure costs become part of process economics.
As a result, a technically successful AI deployment can still be an economically poor one.
Institutions need visibility into authority, access, accountability, performance and cost throughout the lifecycle of AI-enabled work.
Leaders will need a common view of AI unit economics, not simply the cost of accessing a model, but the cost of completing a workflow.
The institutions that understand this early will be better positioned to decide where autonomy creates genuine operating leverage and where traditional automation remains the better answer.
Kyndryl's 2026 People Readiness Report found that organizations are rapidly expanding AI adoption while simultaneously confronting the challenge of redesigning work, developing skills and establishing governance around human-machine collaboration. In banking, those pressures may become more pronounced because agentic AI is more than a productivity tool. It redistributes work.
Some activities will remain human. Others will be collaborative. Still others may become largely machine-executed, with people managing policy, exceptions and accountability.
Before scaling agentic AI, banking leaders should ask these questions: Can an AI-enabled workflow operate safely across organizational boundaries? Can institutions trace consequential decisions, govern agent actions, measure machine-executed work and ensure humans can intervene when needed, all without creating another layer of complexity?
The answers will reveal more about the organizational readiness than the number of pilots completed or agents deployed.
Large and small institutions will approach the transition differently, but scale alone will not determine the outcome. What will matter is whether an institution can convert AI investment into a coherent operating capability.
The first phase of enterprise AI was about capability: What can the technology do? The next phase is about architecture and accountability: How should an institution organize itself when intelligence is distributed across people and machines?
Intelligence is becoming more abundant. Coordination is becoming more important.
The institutions that create advantage in the next decade will not simply deploy intelligence at scale. They will coordinate it.
