Banks Put AI Agents Through Performance Reviews

Highlights

BNY’s open payment investigations fell nearly 80%.

DBS turned its best bankers’ know-how into software.

Brokerage regulators want agent decisions on the record.

A fund manager in Hong Kong emails BNY at 2 a.m. New York time about a transfer that hasn’t settled. A software agent can read it, pull the records and draft a reply. A human manager reviews the draft before it goes out, Microsoft reported in a case study published in May.

    Get the Full Story

    Complete the form to unlock this article and enjoy unlimited free access to all PYMNTS content — no additional logins required.

    Subscribe to our daily newsletter, PYMNTS Today.

    By completing this form, you agree to receive marketing communications from PYMNTS and to the sharing of your information with our sponsor, if applicable, in accordance with our Privacy Policy and Terms and Conditions.

    That agent is one of about 140 digital employees at BNY, PYMNTS reported in April. Each one has login credentials, an employee number in the corporate directory and a supervisor who assigns tasks and reviews the work.

    Getting from an idea to a working agent takes 16 steps of review boards, compliance gates and model risk evaluations at BNY, Microsoft said. The bank moves $2.5 trillion in transactions a day. A misrouted payment can freeze a billion-dollar transfer.

    BNY Runs Its AI Agents Through Performance Reviews

    Managers track each digital employee on a live dashboard, which shows processed volume, cycle times, missed transactions and approval outcomes. When reviewers kept sending back the payments agent’s responses, its team coached and retrained it the way it would a new hire. The pushback came from reviewers’ old habits, not from wrong answers.

    That payments agent replaced scripted bots that kept breaking. Maintaining them took so much effort that “the humans were effectively doing two jobs instead of one,” Rachel Lewis, who leads AI enablement for BNY’s operations, told Microsoft.

    The agent got one job. It reads the vendor address on a cross-border payment, confirms the country code and passes the payment to a person for review. Validations that took up to six minutes now clear in under 30 seconds, and open payment investigations fell nearly 80%, Lewis said. Both figures are the bank’s own and unaudited.

    Michelle O’Reilly, BNY’s head of talent, leads a team building a Digital Employee Agency. It decides when a job goes to a person, a contractor or an agent. It will also cover agent onboarding, scorecards and eventually offboarding.

    We’d love to be your preferred source for news.

    Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!

    DBS Hands First Drafts of Credit Memos to Agents

    Singapore-based DBS gave credit agents to about 1,500 employees globally in August, after a 150-person pilot, the bank announced at the time. The agents handle more than 70 tasks. They pull from annual reports, industry research and internal records to build a first draft of a credit memo for a corporate client.

    Credit memos and related work can take up to 40% of a relationship manager’s time, DBS said. The bank aims to cut that time by at least 30%. That figure is a goal, not a result. Bankers work with the agents to turn the draft into the final memo.

    The tool captures “the knowledge and insight of our best relationship managers and credit risk managers,” Han Kwee Juan, DBS’s group head of institutional banking, said.

    FINRA Asks Firms to Trace What Agents Do

    FINRA’s 2026 Annual Regulatory Oversight report, released in December for the brokerage firms it oversees, added a new section on generative AI. Part of it covers AI agents. The report flags agents acting autonomously without human validation and approval. It also flags agents acting beyond a user’s scope and authority.

    Auditability is another risk on the list. Multistep agent reasoning can make outcomes hard to trace or explain, the report said. FINRA suggests firms consider where to place human oversight and how to track agent actions and decisions. It also points to guardrails that limit what an agent can do.

    State examiners are adding their own questions. The Conference of State Bank Supervisors released an AI supervisory framework on Sept. 16, PYMNTS reported. Each state regulator decides how much of it to use.