AI agents for compliance teams in retail banking in Singapore

Scam losses, fair dealing and sales to retail investors dominate retail banking compliance in Singapore. AI agents can test scam controls against what the bank promised, review sales files for vulnerable customers, check marketing against the PDPA's consent and Do Not Call duties, and draft findings. Compliance officers decide outcomes. MAS, the FEAT principles and the PDPA set the rules.

What fills a Singapore retail bank compliance team's week?

Scams take the largest share. Customers lose money to impersonation and phishing, and the bank must show that its cooling-off periods, transaction alerts, kill switches and reporting channels worked as designed. When a customer claims the bank failed in its duties, compliance checks whether each control fired, and when.

The second load is fair dealing. Singapore banks sell structured deposits, unit trusts and insurance through branch staff and relationship managers. Compliance reviews whether the product suited the customer, whether older or less educated customers were given the extra safeguards the bank's policy requires, and whether staff pay rewarded good conduct rather than volume.

The rest of the week goes on complaints, data protection queries, marketing reviews, and evidence for MAS inspections.

Where can AI agents help?

Four uses fit, each certified separately:

• Scam control testing (depth 1). For a disputed scam case, the agent reads the transaction log, alert history, login and device records, and the bank's own control standards, and builds a timeline showing which control fired and which did not. It cites each record. The compliance officer decides whether the bank met its obligations.
• Sales file review (depth 1). The agent reads the customer's profile, the product's risk rating, the call recording transcript and the signed forms, and flags gaps: a missing assessment of knowledge, a vulnerable customer sold a complex product without a second-person review, a recording that does not match the form.
• Marketing and consent checks (depth 1). Before a campaign goes out, the agent checks the target list against consent records and the Do Not Call status of each number, and flags anyone the bank may not contact.
• Inspection evidence packs (depth 2). The agent drafts the evidence pack MAS has asked for, linking each statement to a sealed record from the system that holds it. A person reviews and sends.

Sources matter. A transaction log from the core banking system is a sealed record. A customer's own account of a scam call is declared, and a screenshot of a phishing message is external. The agent must keep those classes apart in its timeline, because a finding that rests on an unverified claim will not survive challenge. The guide to AI compliance agents as the second line, never the last word explains why that distinction drives every compliance use.

What stays with people?

Whether the bank compensates a scam victim is a decision for a named officer. It turns on judgement about what the customer knew and what the bank should have done, and the outcome is felt by a real person who has lost money.

Findings of unsuitable sales, staff disciplinary referrals, and breach reports to MAS also stay human. An agent can show that a form was signed before the recording ended; a person decides whether that was a training lapse or misconduct.

Data subject responses stay with a person too. The agent can assemble what the bank holds, but a person decides what is disclosed and what is withheld under the PDPA's exceptions.

Which rules apply in Singapore?

MAS supervises retail banks for conduct, prudential soundness and technology risk. Its expectations on fair dealing require banks to treat retail customers fairly, sell suitable products and handle complaints properly. Its AML/CFT rules for banks, which follow the FATF Recommendations, cover customer due diligence and suspicious transaction reporting. Singapore also has a framework sharing scam-loss responsibility between banks and telecoms operators where specified duties are not met.

For AI, three frameworks matter. The FEAT principles (fairness, ethics, accountability, transparency) set MAS's expectations for AI and data analytics in financial services. The Technology Risk Management Guidelines apply to the agent and its supplier as they would to any critical system: governance, testing, change control and outsourcing. IMDA's Model AI Governance Framework adds practical guidance on human involvement and accountability.

The Personal Data Protection Act 2012, enforced by the PDPC, governs the customer data the agent reads, including its Do Not Call provisions for marketing. Our guide to the Singapore PDPA and AI agents covers consent, purpose and transfers. Where customers share data from other institutions through APIs, the consent behind that flow is part of the review; see what open banking is. For the wider picture, read how AI agents are regulated in Singapore.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

How do you start in the first 90 days?

Days 1 to 30: pick scam control testing. Write down the bank's control standards as the agent's sources and name the officer who owns scam reimbursement decisions. Apply FEAT to the agent's role and pass it through your technology risk process like any new system.

Days 31 to 60: replay the agent against fifty closed scam cases, including ones the bank compensated and ones it did not. Compare its timelines with the officers' files. Every place the agent misread a log, or treated a customer's statement as fact, becomes a test case.

Days 61 to 90: add sales file review on a sample of structured deposit sales to customers over a set age. Keep both uses at depth 1 until the certificate shows the agent holds its floor on the bank's own cases, then allow it to draft inspection packs at depth 2.

Where Pryme fits

In Pryme, the scam timeline agent reads sealed records from the core ledger and alerting system and labels the customer's own account as declared. Attest replays it against your closed scam cases before it reaches depth 1 in production, and every flag it raises is filed so a reviewer can rebuild how a finding was reached.