The most important question about agentic AI in banking may not be how intelligent the agents become, but how close banks are prepared to let them get to the systems that actually move money. Giving an AI assistant access to policies, documents or customer conversations is one thing; allowing an agent to interact with accounts, payments, lending decisions and core banking workflows is something considerably more consequential. Yet that is increasingly where the technology is heading. As AI shifts from answering questions to executing tasks, the competitive advantage may no longer belong to banks with the most sophisticated models. It may belong to those whose underlying infrastructure can safely give intelligence the context, permissions and real-time access required to act.
AI is moving towards the banking core
The first phase of generative AI adoption in banking largely happened around the edges of the institution. Banks introduced employee copilots, customer-service assistants, document summarisation tools, coding support and systems capable of helping relationship managers retrieve information more quickly.
These applications could deliver productivity improvements without fundamentally changing the architecture responsible for accounts and transactions. The AI could recommend an action, draft a response or identify information, while established banking systems and employees remained responsible for execution.
Agentic AI begins to challenge that separation.
An agent designed to complete an entire banking process needs more than access to a language model. It requires accurate customer and account information, knowledge of bank policies, awareness of transaction history and the ability to interact with multiple systems. In some circumstances, it may also require permission to initiate an action.
That brings AI much closer to the core banking environment.
The difference is significant. An assistant might tell an employee that a customer qualifies for a particular product. An agent could potentially verify eligibility, assemble the required information, initiate the workflow, obtain approval and update the relevant systems.
The intelligence is only one part of that process. The infrastructure underneath it determines whether the action can actually happen.
Legacy architecture could become an AI constraint
Banks have spent years discussing the limitations of legacy technology in terms of cost, product development and customer experience. Agentic AI introduces another dimension to the modernisation debate.
Many banking environments contain decades of systems connected through APIs, middleware, integration layers and manual processes. Data may exist across several platforms, while different applications maintain their own business rules and permissions.
That complexity is manageable when humans navigate between systems or applications operate according to predetermined workflows. It becomes more difficult when an autonomous agent needs to understand context and complete a process across several systems in real time.
A bank could therefore have access to highly capable AI models while still being unable to use them meaningfully across core operations.
This is where the relationship between AI strategy and banking infrastructure is likely to become much closer. The question will increasingly move from “Which AI model should we use?” towards “Can our architecture safely allow AI to do anything useful?”
For institutions still dependent on fragmented systems and slow integration cycles, the answer may be uncomfortable.
The core could become an intelligence platform
Modern core banking platforms have largely competed around flexibility. Cloud deployment, APIs, composability and faster product configuration have become important selling points as banks try to move away from monolithic architectures.
AI could change what banks expect from the core again.
Instead of functioning primarily as a reliable system of record, future core platforms may increasingly become environments through which intelligent systems can obtain context and execute controlled actions.
That does not mean putting a large language model in charge of the ledger. It means designing infrastructure so that authorised AI systems can securely retrieve information, understand the relevant business rules and perform narrowly defined tasks without requiring another collection of bespoke integrations.
Standards such as Model Context Protocol could become relevant here because they offer a common mechanism through which AI applications can access tools and contextual information. Whether MCP itself becomes dominant is less important than the architectural direction it represents: AI needs structured ways to interact with enterprise systems.
Banks that solve this connectivity problem could move significantly faster when deploying new agents because each application would not need an entirely new integration framework.
Autonomy makes governance part of the architecture
Giving AI access to banking infrastructure also changes the meaning of AI governance.
When a model generates an inaccurate internal summary, the consequences may be manageable. When an autonomous system can initiate a payment, alter a workflow or trigger a customer action, the acceptable margin for error becomes much smaller.
The controls therefore need to exist before the agent acts.
Banks will need granular permissions defining what an agent can see, which actions it can initiate, the monetary or risk thresholds within which it can operate and when human approval becomes mandatory. Every significant action will also need to be traceable so that the institution can determine what information the agent used, what decision it reached and why an action occurred.
This could make AI governance less of a standalone compliance exercise and more of an infrastructure requirement.
An autonomous banking environment without strong identity, permissions, auditability and human escalation mechanisms would introduce risks that few regulated institutions could accept. The banks that progress fastest may consequently be those that treat governance as an enabler of autonomy rather than an obstacle to it.
Banks may need to redesign processes before automating them
There is another complication. Connecting AI agents to existing systems does not automatically create an intelligent bank.
Many banking processes contain unnecessary approvals, duplicate data entry, manual workarounds and procedures created to compensate for limitations elsewhere in the organisation. Giving an AI agent responsibility for navigating those processes could automate complexity rather than remove it.
Banks may therefore need to reconsider workflows before deciding where autonomy should be introduced.
A well-designed agentic process might collapse several stages into one continuous workflow because the agent can gather information, validate conditions and coordinate systems simultaneously. That opportunity disappears if institutions simply replicate every existing human step digitally.
This is why AI agents in banking could ultimately become as much an operating-model discussion as a technology discussion. The greatest gains may come from redesigning how work is performed rather than inserting agents into processes that were built for another era.
Core banking decisions are becoming AI decisions
For technology leaders, this convergence creates a longer-term strategic question.
A core banking platform selected today could remain inside an institution for a decade or considerably longer. Banks evaluating modernisation programmes therefore need to consider not only the products and transaction volumes they expect to support, but also the type of intelligent systems that may interact with that infrastructure in the future.
The ability to expose trusted real-time data, apply granular permissions and allow controlled machine-to-system interaction could become increasingly important selection criteria.
This could also reshape competition among core banking providers. Cloud-native architecture and open APIs will remain important, but banks may begin asking different questions: How easily can AI agents understand the platform? What actions can they perform? How are permissions managed? Can every autonomous decision be reconstructed? How quickly can a new agent be connected without creating another integration project?
These questions would have sounded premature only a few years ago. They are becoming increasingly relevant as AI moves from experimentation towards execution.
What it means for the industry
- AI and core modernisation are converging: Banks may find it increasingly difficult to develop an AI strategy independently of decisions about their underlying banking architecture.
- Infrastructure could matter more than model choice: Access to powerful AI models will become increasingly widespread, making secure access to trusted data and operational systems a more meaningful differentiator.
- Core platforms may evolve beyond systems of record: The next generation of banking infrastructure could also serve as a controlled execution environment for intelligent agents.
- Governance will determine how far autonomy can go: Permissions, thresholds, audit trails and human intervention will need to be designed into agentic banking from the beginning.
- Process redesign will be critical: Banks that simply automate existing complexity may capture far less value than those willing to rethink workflows around what intelligent systems can actually do.
