AI agent training for loan officers in retail banking in Singapore

Every loan officer at a Singapore retail bank who approves an agent's draft should know how the agent reached it, which sources it trusted and what the officer is accountable for. Agents can compute debt servicing ratios, verify income and draft offers. Officers keep exceptions, declines, hardship and disbursement, and MAS's FEAT principles expect that accountability to be clear.

What do retail loan officers in Singapore work on?

Property loans are the core. Each application is tested against MAS rules that limit how much of a borrower's income may go to debt repayments and how much may be lent against a property, with different treatment for public housing and private property and for second and subsequent loans. Buyers often draw on their national retirement savings towards the purchase, and the officer must account for that in the file.

Unsecured lending, from personal loans to credit card limits, has its own income-based limits and borrowing caps. Officers deal with applicants close to those limits and with existing customers asking for more.

The applicant pool is diverse. Many borrowers are foreign professionals on employment passes, with income in several currencies or a mix of salary, bonus and commission. Variable income needs careful treatment under the rules on how much of it can be counted.

Which tasks can an agent take, at which depth?

• Ratio calculations (depth 1). The agent calculates the debt servicing ratio and loan-to-value for the application, using the bank's sealed ledger data for existing loans and the credit bureau report, which is signed.
• Income verification (depth 1). The agent compares declared income with tax assessments and salary credits, applies the rules on variable income and flags differences.
• Offer drafting (depth 2). The agent drafts the offer letter with terms and conditions for an officer to approve and send.
• Limit reviews for existing customers (depth 3). Within the bank's bounds and the regulatory caps, the agent can process routine limit reviews for customers with stable histories. Anything near a cap is held.
• Disbursement (depth 4). Held for a named officer.

Each verdict names the bound that decided. A case held because the applicant's bonus income could not be confirmed from a signed source is different from one held because the agent's certificate does not cover foreign-currency income, and officers should learn to tell them apart.

What stays with the loan officer?

Exceptions to bank policy, declines and the reasons given, any case close to a regulatory limit where judgement matters, hardship arrangements and disbursement. Officers also keep the decision on whether a borrower's situation, beyond the numbers, makes the loan sensible. See what loan officers should never let an AI agent do.

The numbers can be calculated; accountability cannot. The accountability principle in FEAT asks a bank to be able to say who answers for an AI-supported decision, which is the case for keeping officers as the real control in an AI agent deployment. It is also why the role persists, as argued in will AI agents replace loan officers?.

What do officers need to learn?

Officers need fluency in the vocabulary of governed agents: depth, the four verdicts, the minimum chain of five bounds and provenance classes. They need a review routine for drafted offers that checks the ratios, the income treatment and the terms. They need to know how to correct a draft so the bank can improve the agent's brief and sources, as described in teaching loan officers to bank corrections.

They also need to understand certification. The agent earns depth by being replayed on the bank's own past applications; its certificate lapses after 90 days, and a run that clears no depth drops it to depth 1 the same day. Property market changes often bring rule changes, and keeping the agent's sources current matters after each one.

Which Singapore rules shape oversight?

The Monetary Authority of Singapore (MAS) supervises banks and sets the property lending and unsecured credit rules described above. The MAS FEAT principles cover fairness, ethics, accountability and transparency in the use of AI and data analytics: a bank should be able to justify its use of an agent in credit, show it treats customers fairly and explain decisions. The MAS Technology Risk Management Guidelines cover how the bank manages technology risk, including systems provided by third parties.

The Personal Data Protection Act 2012, enforced by the PDPC, governs personal data in lending files. IMDA's Model AI Governance Framework offers voluntary guidance on human involvement in AI-supported decisions, which aligns with keeping officers in the loop on credit.

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

How should the first 90 days run?

Days 1 to 30. Officers, team leads, the credit risk owner and branch managers complete Foundations (about 4 hours). The agent calculates ratios at depth 1 while officers do the same by hand for a sample of files.

Days 31 to 60. Officers take the Operator modules on how a governed agent decides, human review, keeping knowledge current and when something goes wrong. Offer drafts move to depth 2. Leads review every case held near a regulatory limit.

Days 61 to 90. Officers sit the Operator exam (40 questions in 60 minutes; 32 correct to pass) and complete the two-week practicum. Log each review decision in the practicum log template for loan officers for the assessor.

See AI agent skills by role for the other roles involved.

Where the Academy fits

Pryme Intelligence Academy certifies officers through a written exam and a two-week practicum of real review decisions, signed off by a Pryme assessor. Certified Pryme Operator is valid for 12 months, and a public register lists certified people by name, credential and expiry. That gives a bank a clear answer when asked, under FEAT, who is accountable for agent-drafted credit decisions.