As financial institutions expand their use of autonomous AI, controlling access to sensitive data across fragmented systems is becoming a critical security and operational challenge.
Imagine an AI agent assisting a bank's fraud investigation team. To investigate suspicious activity, it retrieves transaction histories, reviews customer records, queries risk databases and identifies connections between accounts. Each step may be appropriate for the investigation. But what happens if the agent accesses information beyond its authorized scope, encounters conflicting security policies or initiates an action that requires human approval?
These are becoming increasingly important questions as banks explore AI systems capable of executing complex workflows with greater independence.
Recent developments demonstrate the urgency. In October, major U.S. banks began restricting third-party AI agents from accessing customer accounts amid concerns about unauthorized activity, security and accountability. Axios reported that some institutions had detected agents accessing accounts using customer credentials, making it difficult to distinguish automated activity from actions initiated directly by customers.
While these concerns involve external agents acting on behalf of consumers, banks face related challenges as they deploy autonomous AI within their own operations.
Financial institutions have established extensive controls over how employees and applications access sensitive information. AI agents introduce additional complexity because they can dynamically select data sources, use tools and execute multiple steps to complete a task.
For banks operating across legacy systems, cloud platforms and specialized applications, ensuring that these activities remain authorized, visible and auditable presents a significant challenge.
When AI agents operate beyond traditional access models
Traditional banking applications generally operate within predefined workflows and permissions. AI agents can make decisions about which information to retrieve and which tools to use as they pursue an assigned objective.
Consider an agent supporting anti-money laundering investigations. It might access customer onboarding information, examine transaction histories, retrieve sanctions screening results and review previous investigations to identify potentially suspicious activity.
The agent may need information from several systems, each governed by different access policies. Its ability to retrieve information from one source does not necessarily mean it should have permission to access related records elsewhere.
This raises several questions for banks.
How can an institution verify that an agent is accessing information for an authorized purpose? Should the agent inherit all the permissions of the employee who initiated the task? How can the bank track which data influenced an agent's recommendations? And where should human approval be required before an agent takes action?
These questions become more difficult when agents interact with multiple applications or invoke other automated services.
Giving an agent broad access to simplify development may accelerate an initial deployment, but it can also increase the potential consequences of mistakes, compromised credentials or unintended behavior.
Banks need to establish clear boundaries around agent identities, data permissions, tool access and permitted actions. They also need sufficient visibility to determine whether those boundaries are being respected.
Fragmented banking data adds another layer of complexity
Banking technology environments reflect decades of investment, acquisitions, regulatory requirements and modernization initiatives.
Customer information may reside in core banking platforms, transaction records in payment systems, risk information in specialized applications and historical data in warehouses or data lakes. Increasingly, this information spans both on-premises infrastructure and multiple cloud environments.
Each system may have its own access controls, data classifications and governance policies.
For traditional analytics, this fragmentation can create delays and additional integration work. For AI agents that dynamically retrieve information across systems, it can also create gaps in oversight.
Consider an agent investigating potential fraud involving several related accounts. To understand the activity, it may need to combine payment records, account ownership information, customer interactions and historical alerts.
Even if the agent is authorized to perform the investigation, it should not automatically receive unrestricted access to every underlying dataset. Some records may contain sensitive personal information, while others may be subject to different retention, privacy or jurisdictional requirements.
Banks must ensure that the agent receives the information necessary for its task while maintaining the restrictions that apply to each source.
Data duplication can complicate this process. When information is copied into multiple environments, changes to access policies, classifications or underlying records may not be reflected consistently across every copy. Banks also need to consider data freshness and lineage when agents rely on replicated information.
These challenges make data architecture an important part of agentic AI governance.
Three priorities for governing AI access to banking data
As banks evaluate agentic AI, three data infrastructure capabilities deserve particular attention.
1. Apply consistent access controls across distributed environments.
Banks need a reliable way to enforce appropriate data permissions regardless of where information resides.
Role-based and attribute-based access controls, together with row-level filtering and column masking, can help restrict exposure to sensitive information. Agent identities should be explicitly defined, with permissions appropriate to their assigned responsibilities rather than unrestricted access inherited from human users.
A consistent policy framework can reduce the risk of an agent encountering different authorization rules as it moves between systems.
However, banks must also account for controls that remain specific to individual platforms, applications and jurisdictions. A common access layer should complement those safeguards rather than bypass them.
2. Connect data access activity to agent-level accountability.
Banks need to know more than whether an agent successfully completed a task.
They should be able to determine which information the agent accessed, when it was retrieved, what permissions were applied and which subsequent actions were taken.
Query logging, data lineage and monitoring can help establish visibility into information access. Connecting those records with agent execution logs can provide a more complete account of how an agent arrived at a recommendation or initiated a workflow.
This visibility is particularly important in regulated processes such as fraud investigations, credit operations and compliance reviews, where decisions may require examination or explanation.
Data access logs alone cannot fully explain an AI model's reasoning. But they can provide essential evidence about the information used and the controls applied.
3. Enable governed access without unnecessary data movement.
Banks do not necessarily need to consolidate every relevant dataset into a single repository to support AI agents.
A federated data access approach can allow authorized applications to query information across multiple environments while reducing the need to duplicate data for every new use case.
A common access layer can help institutions standardize aspects of authorization, monitoring and data retrieval while preserving investments in existing systems.
This approach can be particularly valuable when agents need information from several business units or technology platforms.
Federation is not appropriate for every workload. Some applications require streaming data, local processing, caching or other architectures to meet performance and availability requirements. Banks must evaluate the appropriate approach based on the needs of each application.
The objective is to make authorized data available to AI systems without creating unnecessary complexity or weakening existing safeguards.
Data governance is only one part of controlling autonomous AI agents
A well-governed data environment can help limit what an AI agent is permitted to see. Banks must also control what an agent is allowed to do with that information.
An agent authorized to analyze transactions, for example, should not automatically have permission to initiate payments, modify customer records or change account restrictions.
Those actions require separate controls governing tool permissions, transaction authority and approval workflows.
Banks should also consider how they will respond when an agent behaves unexpectedly. Can its access be revoked immediately? Can an incomplete workflow be stopped? Is there sufficient evidence to reconstruct what occurred? Can affected transactions or records be identified for review?
These operational safeguards become increasingly important as agents move from assisting employees to executing more complex activities.
Banks also face different governance challenges depending on whether AI agents operate internally or on behalf of customers. Third-party agents accessing consumer accounts raise questions about customer consent, identity verification and liability. Internally deployed agents introduce responsibilities around enterprise permissions, workflow execution and oversight.
Both require clearly defined authority and accountability, but the mechanisms for managing them may differ.
Building confidence in agentic AI
AI agents offer opportunities to accelerate fraud investigations, improve customer service, support compliance teams and reduce repetitive operational work.
Realizing those benefits at scale will require banks to establish confidence in how agents access information and execute tasks.
For banking technology and risk leaders, several questions should guide deployment decisions:
- Can the institution identify every agent accessing sensitive information and verify its permissions?
- Are data access policies enforced consistently across the environments an agent uses?
- Can the bank reconstruct which information an agent accessed and what actions followed?
- Are consequential actions subject to appropriate authorization and human oversight?
- Can the institution contain or revoke an agent's access when unexpected behavior occurs?
The answers depend on coordinated controls across data infrastructure, identity management, application security and operational risk management.
As financial institutions expand agentic AI adoption, fragmented data environments and inconsistent governance can limit their ability to deploy autonomous systems responsibly.
Banks have spent decades building controls around access to sensitive financial information. As AI agents take on more responsibility, those controls will face new tests, particularly where data remains scattered across systems with different policies and permissions.
Article top image credit: iStock.com/JuSun