Who controls the AI agents transforming financial sector businesses?

by Steven Jones, CTO, Kocho

When an AI agent moves client money in financial services, who gave it the authority to act? It is a question that financial institutions need to answer as regulators and insurers become more concerned about the potential risks.

Across the financial sector, AI agents are ready to deliver major efficiency gains. AI is moving from conducting analysis and recommending next steps to acting autonomously. But the question of the identity controlling the agent is worrying regulators.

The Financial Conduct Authority’s (FCA) view is that controls over access and identity designed for human users are no longer sufficient. In October 2026, Jessica Rusu, the FCA’s chief data, information and intelligence officer confirmed it plans a framework for AI identity, delegated authority, accountability and execution.

How agentic AI changes the security equation

The FCA is building on the forward-looking Mills Review of July 2026, and expects AI agents to carry out more work independently, moving across connected systems, using different levels of access as they complete tasks.

Access creates risk. An agent helping a client transfer money between platforms needs to read account information to analyse the options and prepare a transaction. Each step involves permissions which create risks if they remain active after a task is completed.

These are risks frequently inherited from accumulated service accounts, outdated entitlements and poorly governed third-party access that all goes unmonitored. If attackers are able to acquire credentials and control such agents, they can quickly move through connected systems.

Steven Jones

Why Zero Trust should extend to AI identities

To cut down these risks, institutions should extend Zero Trust security to AI agents as well as people. Every digital identity capable of accessing data or triggering an action needs a defined purpose, an accountable owner and a clearly controlled lifecycle. No identity should receive wide-reaching permissions just because it operates in a trusted environment.

Agents differ from most digital identities because their access need changes with the task. Rather than granting standing privileges, firms should only approve a specific action and ensure approval ends once the task is complete.

Financial firms also need to separate preparing an action from authorising it. An agent might assess investment options or assemble a payment instruction, but a high-value transaction should require active approval from an authorised person whose identity is verified.

To reduce opportunities for fraud and theft, controls must follow the agent across cloud, internal and third-party platforms. Active record-keeping is also essential because effective Zero Trust requires an organisation to have detailed knowledge of where its sensitive data resides, which identities can access it and whether those permissions are justified.

Who watches the agents?

Organisations need to know who initiated a task, the instructions provided and why the agent took specific actions. When several systems and providers are involved, this becomes that much harder. Rusu warned that systems judged safe at deployment may behave differently under stress, at scale or following changes to underlying models, highlighting the need for continuous, conditional assurance.

Organisations need to record agent activity and their instructions so teams can effectively investigate suspicious activity and explain decisions to regulators. This is not just important for relationships with regulators – clients, insurers, and auditors all want to see that access to sensitive systems and information is fully under control. The EU’s AI Act is also introducing stricter, risk-based AI governance requirements that nobody can ignore.

Resilience

The corollary of this increased focus on agents is that identity governance should be part of risk management, compliance and operational resilience in a coordinated control model. The objective should be to apply safeguards proportionate to the consequences of each action, reserving additional checks for costly, high-risk decisions.

Operational resilience is equally important, including stress-testing of AI systems and maintenance of fallback options. Automation increases the potential speed and scale of incidents, particularly across interconnected systems. Institutions need to know what happens when instructions are manipulated, access is denied or an agent behaves unexpectedly. Clear escalation procedures, the ability to halt automated activity and reliable fallback arrangements are essential. This is where Zero Trust will limit the potential for damage.

As agents take on greater responsibility, static, complacent approaches to identity become very risky. Financial sector firms need time-limited access backed by a form of telemetry to provide assurance and human oversight.

Organisations are about to become far more accountable for the outcomes of their AI agents and must act to protect themselves. AI agents will transform efficiency but must be fully under scrutiny if they are not to create new risks.

spot_img
spot_img

Subscribe to our Newsletter