
As AI agents increasingly operate independently within sensitive financial systems, institutions must reassess their approaches to identity, permissions, and accountability. The Financial Conduct Authority’s goal of becoming an agentic regulator reflects a broader shift in the financial services sector. With AI transitioning from analyzing data and suggesting actions to executing tasks autonomously, banks and other institutions can enhance efficiency and customer service, but they must also address the key issue of authorization and access control.
Establishing Agent Identity
In October 2026, Jessica Rusu, the FCA’s chief data, information, and intelligence officer, announced plans to develop a framework for agent identity, delegated authority, accountability, and execution. This move acknowledges the need for more robust controls over access and identity for AI agents than those designed for human users.
Building on the Mills Review from July 2026, which examined the impact of AI on retail financial services up to 2030, the FCA anticipates that AI agents will perform more tasks independently, handling connected systems and utilizing varying levels of access. An agent assisting a customer with transferring funds between financial platforms must access account information, evaluate options, and prepare the transaction, each step involving its own set of permissions and risks, particularly if access remains active after task completion.
The integration of AI agents into existing systems can expose hidden vulnerabilities, as financial organizations often accumulate outdated service accounts, poorly managed third-party access, and uncontrolled entitlements. If attackers gain access to an AI agent or its credentials, they can rapidly move through connected systems, executing malicious activities.
Read Also: Corporate Travel Policies Miss Hidden Time Costs
Microsoft’s Secure access in the age of AI report emphasizes the growing problem of fragmented identity and network access controls, with AI agents expanding the number of identities, access points, and potential vulnerabilities, thus increasing the attack surface. Implementing Zero Trust principles, which verify access at every step, is essential for both human and AI identities, ensuring that each digital identity has a defined purpose, an accountable owner, and a controlled lifecycle, without receiving broad permissions solely due to operating in a trusted environment.
Unlike traditional digital identities with relatively fixed permissions, AI agents require dynamic access that changes based on the task. Instead of granting standing privileges, institutions should adopt a temporary access approach, approving specific actions only when necessary and revoking permissions upon task completion, thereby reducing the risk of unnecessary access to sensitive systems.
Effective Zero Trust also demands detailed knowledge of sensitive data locations, accessible identities, and justified permissions, aiming to prevent autonomous systems from gaining excessive access without compromising efficiency benefits. Distinguishing between action preparation and authorization is key, as agents may be allowed to assess investment options or assemble payment instructions but should require additional security layers for high-value transactions, ensuring that greater autonomy does not compromise human accountability or facilitate fraud.
For identity governance to be effective, it must encompass the agent’s entire activity, not just individual systems. As the regulatory environment evolves, institutions must ensure their governance frameworks demonstrate compliance with emerging standards, including maintaining clear audit trails that connect actions to authorized identities, regardless of automation complexity. This visibility is essential for operational resilience, as institutions risk falling out of step with internal risk management and external regulatory expectations without it.
The Challenge of Accountability
The rise of AI agents poses accountability challenges, as organizations need to track who initiated tasks, what instructions were given, and why agents took specific actions. The Mills Review highlights the difficulty in demonstrating how outcomes were reached as decisions are delegated to AI, especially with multiple systems and providers involved. Rusu warns that systems considered safe at deployment may behave differently under stress or after changes to underlying models, emphasizing the need for continuous assurance. Detailed monitoring of AI agents, with up-to-date records of their activities, is necessary to investigate suspicious activity, explain decisions to regulators, and maintain oversight as AI systems become more autonomous.
Read Also: Kazakhstan Probes Manka Gold Deposit Fraud Allegations
Organizations must track the instructions agents receive, the systems they access, and the actions they take, facilitating the reconstruction of the full sequence of events to ensure every action can be traced back to an authorized instruction. Without such records, distinguishing between genuine errors and deliberate attempts to subvert the system becomes challenging. The regulator expects financial institutions to maintain evidence of the full sequence of events, showing the importance of accountability in AI governance.
Beyond Security to Governance
Bringing identity, access management, data protection, and security operations into a coordinated control model provides firms with a clearer understanding of excessive privileges and emerging risks. The objective is to apply safeguards proportionate to the consequences of each action, allowing routine activities to proceed efficiently while reserving additional checks for high-risk decisions.
Operational resilience is equally key, with Rusu urging firms to stress-test AI systems, plan for failure, and maintain fallback options, understanding dependencies on external providers. Automation accelerates processes but also increases the potential speed and scale of incidents, particularly across interconnected systems, necessitating clear escalation procedures, the ability to halt automated activity, and reliable fallback arrangements, with Zero Trust controls limiting potential damage by restricting agent access and changes before intervention.
Implementing Full Governance and Resilience
Financial institutions must integrate identity governance into risk management and compliance strategies. Organizations need to monitor the entire activity of AI agents, including the instructions they receive and the actions they take.