Anti-scam operations dominate Singapore retail banking, and AI agents can carry much of the casework: emergency freeze follow-ups, cooling-off exceptions, scam claim files and digital servicing requests. Unfreezing accounts, lifting protections early, loss decisions and credit stay with people. MAS rules, the Technology Risk Management Guidelines, the FEAT principles and the Personal Data Protection Act 2012 apply.
What do retail banking operations leaders in Singapore spend their time on?
Scam defences. Singapore banks have added layers of protection for retail customers: a self-service option to freeze accounts in an emergency, cooling-off periods after a new device is registered or security settings change, alerts on outgoing transfers, and the option to lock part of savings so it cannot be moved digitally. Each protection creates operational work when customers want it reversed, questioned or explained.
Scam claims. When a customer is scammed, the bank must assess what happened, which protections fired, and whether it met the duties expected of it under the shared responsibility approach. Files are detailed and timelines tight.
Digital servicing. Most retail banking is done on mobile. Servicing requests, such as limit changes, card replacements and account updates, arrive in volume and must be handled quickly without weakening security.
A multilingual, ageing customer base. Many older customers prefer Chinese, Malay or Tamil and are targeted by scammers. Service must be clear in their language.
Where can AI agents help Singapore retail banking teams?
- Emergency freeze follow-up (depth 2). After a customer freezes their account, the agent contacts them through a verified channel, gathers what happened and prepares the file for a person to unfreeze. It never unfreezes.
- Cooling-off exception requests (depth 2). When a customer asks to bypass a cooling-off period, the agent checks the request against the bank's criteria and drafts a recommendation. A person decides.
- Scam claim files (depth 2). The agent builds the timeline from device, login, alert and transfer records, maps each event to the expected duty and drafts the assessment for the scam team.
- Digital servicing requests (depth 3). The agent processes routine requests such as card replacements and address updates within set limits and checks. Limit increases are drafted.
- Multilingual explanations (depth 2 and 3). The agent explains protections and processes in English, Chinese, Malay or Tamil from approved content, certified for each language.
Cooling-off exceptions show why speed is not the goal. Customers asking to skip a protection may be under a scammer's instruction at that very moment. The agent's value is a complete, fast file for the person deciding, not a faster yes.
What stays with people, and why?
Unfreezing accounts and lifting protections early stay with trained staff. Scammers try to persuade customers to request exactly that.
Loss decisions on scam claims stay with people. They allocate money between the customer and the bank and can be disputed.
Limit increases, credit decisions and account closures stay with people.
Suspicious transaction reports stay with compliance, consistent with the FATF Recommendations.
Which rules apply in Singapore?
MAS licenses and supervises banks. Its requirements cover anti-money-laundering, customer protection for unauthorised and mistaken transactions, outsourcing and technology risk. The Technology Risk Management Guidelines set expectations on resilience, access control, change management and incident reporting that reach an agent and its supplier. MAS has worked with the industry on the anti-scam measures described above, and the shared responsibility approach allocates phishing scam losses when expected duties are not met.
The FEAT principles (fairness, ethics, accountability and transparency) guide AI and data analytics in financial services, and IMDA's Model AI Governance Framework sets out voluntary practices on human involvement and operations management; our answer on whether the Model AI Governance Framework covers AI agents explains the fit.
The Personal Data Protection Act 2012, administered by the PDPC, governs customer data; see Singapore's PDPA and AI agents. Data sharing with third parties through bank APIs, discussed in our explainer on what open banking is, follows the same consent and purpose rules. Singapore has no single AI law.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
How should you start in the first 90 days?
Weeks 1 to 4: start with digital servicing requests in shadow. They are high volume, low risk and easy to score.
Weeks 5 to 8: certify routine servicing at depth 3. Shadow scam claim files against cases the scam team closed in the last quarter.
Weeks 9 to 13: add emergency freeze follow-up at depth 2 and multilingual explanations, starting with the language most requested after English. Keep every unfreeze with a person. Follow deploying an AI agent: train, shadow, certify, widen. For a market with stronger hardship and estate workloads, see our brief on operations agents in Australian retail banking.
Where Pryme fits
In Pryme, a request to lift a protection is outside the servicing agent's role scope, so it is refused for the agent and routed to a person. Device and alert logs arrive as sealed sources, and the customer's account of events as declared. Every verdict is filed, which gives the scam team a complete record when a loss is disputed.
Questions people ask
Can the agent unfreeze an account if the customer passes verification?
No. Verification proves identity, not that the customer is free from a scammer's influence. A trained person decides.
Can the agent speak to elderly customers in their language?
Yes, where certified in that language. Keep messages short, use approved content and offer a person on request.
Does the agent need MAS approval?
There is no separate approval for agents. MAS expects you to manage the agent under its technology risk and outsourcing requirements, and to be able to show how.