AI agents in financial services face accountability gap, TechRadar reports

AI agents could reshape how financial services operate, from investigating fraud alerts to processing claims and supporting compliance teams. As AI moves from generating answers to executing actions, firms need to know not just how a decision was reached but what happens next, according to the report. Explainability remains a central question, and it becomes more complex as AI begins, in the words of Financial Conduct Authority Chief Executive Nikhil Rathi, to “coordinate and transact.” The report argues that model explainability is only half the picture. Firms also need operational accountability across the workflow, with agentic business orchestration helping provide visibility into the systems an agent accessed, the actions it took and where human judgement intervened.

The concern is active. According to the TechRadar report, 48% of UK financial services executives say their firms are using agentic AI, yet more than a quarter report having no or limited controls to ensure compliance with laws and regulations. That gap matters because agents create a more complex governance challenge: they can move between systems, interpret information, select tools and determine what happens next. The result can be a black-box journey flow in which decision-making is distributed across multiple tools and systems rather than traceable to clear owners.

The report uses an insurance claim as an example. One agent might collect evidence, compare it against a policy and route the claim for enhanced scrutiny. Another could contact the customer or recommend withholding payment. If challenged, the organization needs the visibility to reconstruct the journey. A small error by an AI agent can be repeated and amplified, the report says. The real risk is not the hallucination itself but what it sets in motion. A single incorrect detail could trigger an account restriction, customer communication or regulatory report before anyone spots the mistake.

That logic reflects the FCA’s operational resilience rules, which expect firms to map important business services, identify vulnerabilities and understand how disruption could travel through them. Agentic AI requires the same visibility into how failures move through an automated journey, according to the report.

Accountability cannot stop at the AI model, the report says. As AI operates across more systems, accountability needs to follow the work. In January, the Treasury Committee warned that the UK’s “wait-and-see” approach to regulating AI in financial services could expose consumers and the wider system to serious harm. It also called for clearer guidance on consumer protection and senior managers’ responsibilities. Firms need controls that work at the level and speed of the process, with clear ownership and an auditable record of the decisions and actions taken.

The challenge is that no single team may have a complete view of an agentic process. Compliance may understand the policy, technology teams the model, operations the customer journey, and a third-party provider part of the infrastructure. An agent can move across all four in seconds. The report describes this as the organizational face of “agent sprawl,” in which agents multiply across systems faster than any single team can track them. A senior manager can therefore remain responsible for an outcome while lacking the evidence needed to understand how it happened.

“Human-in-the-loop” is often used as shorthand for responsible AI, placing a person at a point in the process to review or approve an AI-generated decision. But meaningful oversight depends on what that reviewer can see and whether they can intervene, according to the report. An employee receiving an agent’s recommendation may not know which systems it accessed, whether the information was complete or what actions have already been taken. A credible-looking recommendation can quickly become a rubber stamp.

Controls should reflect the consequence of the action, the report says. Low-risk administrative work may only require monitoring. Restricting a bank account, rejecting an insurance claim or escalating a fraud allegation should demand a much higher threshold for human oversight. The reviewer needs evidence behind the recommendation, a record of earlier steps and the ability to challenge or reverse the action. The goal is not to put a human click in front of every task but to make human judgement count where it is needed.