Express Computer
Home  »  News  »  Banks are giving AI agents the keys. Who checks them before they act?

Banks are giving AI agents the keys. Who checks them before they act?

0 0

For most of the last decade, artificial intelligence in banking played an advisory role. A model scored a transaction, flagged a suspicious payment or estimated a borrower’s risk, and a human decided what happened next. That is changing. A new generation of AI agents can move money, adjust a customer’s credit limit or close a compliance case on their own.

The appeal is speed. The harder question is the one regulators will ask the moment something goes wrong: who checked the agent before it acted, and can the bank prove it?

Gopinathan Rathinavelu, a Phoenix-based data architect with more than 20 years in banking and financial services, believes many institutions are not yet built to answer that question. “Traditional model reviews happen once a year, but an AI agent can change in a week,” he said. “Governance has to run as continuously as the AI does.”

The gap between annual review and weekly change
Bank model governance was designed for a slower world. Models were built, validated, approved and then reviewed on a schedule.

Agentic AI breaks that rhythm. Agents draw on changing data, call other systems and can be updated far more often than traditional models. An agent that behaved safely in March may behave differently in May. A once-a-year review would miss that.

Rathinavelu’s answer is to move governance off the calendar and into the system itself. He sees three essential controls: continuous oversight that watches agents in real time, safeguards that can stop an agent before it carries out a high-risk action, and audit trails detailed and tamper-resistant enough to show a regulator exactly what an agent did and why.

Why it starts with data
“A model can be accurate and still be unusable if you can’t trace its inputs or secure its data,” Rathinavelu said.

That conviction runs through his research. A 2026 study in the International Journal of Computer Information Systems and Industrial Management Applications applies graph neural networks to detect cross-border money laundering in real time. Another paper proposes a scenario-based benchmark for testing AI safety controls in banking. Other work covers explainable AI that automatically produces adverse-action reasons when credit is declined, and automated checks against the EU AI Act and NIST rules.

Before an AI agent can be trusted to act, a bank needs to know its data is sound, its controls hold under pressure, its decisions can be explained and its behavior stays within the rules.

Governance as a team effort
Rathinavelu argues that governing agentic AI cannot belong to a single department. Data architects must make inputs traceable. Risk managers must define what an agent may do and when a human must review. Compliance teams must turn regulations into controls a machine can check. Working in isolation, they leave gaps, and autonomous systems find gaps quickly.

The stakes go beyond fines. A bank that cannot explain why an agent froze a customer’s account risks losing that customer’s trust.

– Rathinavelu is a TOGAF-certified data architect and an IEEE Senior Member, and is the author of AI-Driven Banking Systems: Machine Learning, Fraud Analytics and Blockchain Applications. He received the Knowledge Research Academy’s Best Researcher Award (AREA 2025) and serves on the editorial boards of two international journals. In 2026 he gave keynote addresses at InC4 2026 at CHRIST University in Bengaluru and at ISRC-2026 conference in Malaysia, and he served as a judge at Ulster University’s Agentic AI Hackathon 2026. His work ahead spans AI, AI governance, Agentic AI, GRC and FinTech, all tied to the question this story began with. Banks are handing AI agents the keys. The institutions that succeed will be the ones that can prove, at every moment, who is checking them.

Rathinavelu’s research, books and speaking engagements are undertaken in a personal capacity. The views expressed are his own and do not represent those of any employer.

Leave A Reply

Your email address will not be published.