DBS and Visa pilot bank-controlled AI payments as agents move from advice to action

DBS is piloting AI agents that can pay for customers. Here's what support teams need to prepare

DBS is testing a system with Visa that lets AI agents complete real purchases using bank-controlled credentials. Reports indicate the pilot has already run live transactions, including food and beverage orders on DBS or POSB cards.

This shift moves AI from giving recommendations to taking action. For Customer Support, that means new ticket types, new data to collect, and new playbooks for consent, refunds, and disputes.

How the pilot works (and why it matters to support)

The trial uses Visa Intelligent Commerce to let software agents shop, choose options, and pay. Payment details are tokenized, and the issuing bank runs the approval logic to check identity, limits, and user permissions before money moves.

That keeps the bank in control-but customers will come to you when the "agent did the buying." Your team becomes the front line for clarity, reassurance, and fast resolution.

New conversation types you'll see

• "The bot bought something I didn't want." Customer set broad rules; outcome feels wrong but technically authorized.
• "It exceeded my budget." Agent rule misread or limit not updated; needs rule review and possible goodwill credit.
• "Wrong item/merchant issue." Standard merchant dispute, but initiated by an agent. Route like any fulfillment problem.
• "I didn't approve this." Possible consent gap, shared device, or real fraud. Trigger identity checks and fraud workflow.
• Subscription renewals triggered by agent auto-renew logic. Clarify renewal policy and opt-out path.

Data to capture on first contact

• Exact timestamp, amount, currency, and merchant descriptor
• Last four digits of card and whether a token was used (token reference if visible)
• Agent policy/rule name the customer set (budget, brand, frequency)
• Approval/authorization code and issuer decision notes (if accessible)
• Channel of initiation (app, assistant, smart device, browser extension)
• Device or agent ID, location, and any customer notifications or prompts shown

Suggested triage flow

• 1) Verify identity and confirm the cardholder on the account.
• 2) Check issuer logs: Was the transaction within defined agent rules and spending limits?
• 3) Confirm tokenized credential use and match device/agent ID with the customer's profile.
• 4) Classify: consent issue, fulfillment issue, limit breach, or suspected fraud.
• 5) Take next step: refund/credit path, rule adjustment, merchant dispute, or fraud block and reissue.
• 6) Close the loop: update the agent's rules so the problem doesn't repeat.

Macros and phrasing that reduce friction

• Consent clarity: "I see your agent was allowed to buy within a $X budget for [category]. We can tighten those rules now and review this charge."
• Expectation reset: "Your bank approves each purchase based on your rules and limits. We'll adjust those settings and handle the refund path today."
• Next-step control: "Would you like to pause agent purchases for this category or require a one-tap confirmation first?"

Edge cases to pre-assign

• Partial/duplicate charges: Split tenders or retries by the agent. Validate merchant logs before refund.
• Refunds to tokens: Ensure refunds return to the same token or map correctly if the token rotated.
• Chargebacks: Use standard reason codes (e.g., merchandise/service disputes) and attach agent-rule evidence.
• Shared households: Multiple users, one agent. Require per-user permissions and spending caps.

Metrics to watch

• Agent-initiated dispute rate and refund turnaround time
• Consent confusion rate (tickets where rules weren't understood)
• False positive fraud blocks on agent purchases
• % of cases resolved without a chargeback
• Opt-outs from agent-driven purchases after first incident

Prepare your Help Center

Publish a plain-language page: what agent purchases are, how consent works, how to pause, and how to set limits. Add screenshots of settings and a 3-step "undo" guide.

• FAQ: How do I set a spending cap? How do I require confirmation? How do I see the agent's purchase history?
• Safety: How to freeze agent purchases instantly. What happens to subscriptions during a freeze.
• Disputes: What to expect if an agent misbuys vs. merchant error vs. fraud.

Security and governance context

Analysts point out that customers accept AI suggestions faster than AI decisions with money involved. Visa's model keeps approval logic with the issuing bank and uses tokenization to reduce exposure, which helps trust-but clear consent flows are still your best defense.

If you want deeper context on tokenization, see Visa's overview Visa Token Service. Industry coverage of the pilot has also appeared at Fintech Futures.

What's likely next

Early focus will be low-risk, repeat purchases: groceries, subscriptions, travel bookings, restocks. As comfort grows, expect broader online shopping and bigger baskets-once consent, limits, and refunds feel predictable to customers.

30-day action checklist

• Map agent-purchase touchpoints across your app, devices, and partners.
• Create macros for consent confusion, limit breaches, and merchant disputes.
• Add required case fields for agent ID, rule name, and token flag.
• Set a fast path to Risk/Payments Ops for first-contact resolution.
• Publish a Help Center article with pause/limit instructions.
• Run a mock incident drill: unintended purchase at a known merchant.
• Define goodwill credit guidelines for unclear-consent scenarios.
• Instrument metrics: dispute rate, opt-outs, repeat incidents per user.

Level up your playbooks

For deeper best practices on consent flows, agent telemetry, and dispute handling, explore AI for Customer Support.