An AI assistant can recommend dinner, a flight, or a household purchase without moving money. The moment it completes the transaction, the relationship changes. Convenience becomes authority, and a helpful suggestion becomes a financial action that may later be disputed.
DBS Bank and Visa are testing the infrastructure for that transition through Visa Intelligent Commerce. The pilot allows software agents to search, choose, and pay using credentials issued and controlled by the bank. Real food and beverage purchases have already been processed with DBS or POSB cards.
The significant feature is not that an agent can click a checkout button. It is the permission layer around the action: tokenised payment details, customer-defined rules, and issuer-controlled approval before funds move.
Routine Purchases Provide a Controlled Beginning
Early agent-commerce scenarios include grocery orders, subscription renewals, travel booking, and household restocking. These tasks recur, have familiar merchants or products, and can be constrained by budget, frequency, and preferences.
A customer might allow an agent to reorder a standard item below a fixed amount while requiring confirmation for substitutions or price increases. DBS and Visa plan to extend testing towards broader online shopping and travel, where choices and cancellation terms become more complicated.
Starting with low-risk purchases lets banks measure whether instructions are interpreted correctly and whether customers understand the authority they have delegated. Useful indicators include completed transactions, confirmation requests, rule violations, disputes, refunds, fraud, and how often users revoke an agent’s access.
Adoption is likely to depend less on technical novelty than on predictable behaviour. People may accept automated restocking long before they trust software to choose a costly trip.
Banks can use those early transactions to refine default limits and learn which moments consistently require explicit confirmation.
The Bank Remains the Control Point
Visa’s framework keeps payment credentials tokenised and routes each transaction through approval flows controlled by the issuing bank. The bank can verify identity, enforce spending limits, and decide whether the purchase fits the customer’s permissions.
This positions financial institutions as trust infrastructure for agent-driven commerce. Merchants and agents may handle discovery, but the issuer preserves a final policy boundary around money.
The design also limits exposure of raw card details. An agent should receive only the token and authority required for a specific task, merchant, amount, or period. Permissions need to be visible, easy to change, and automatically expire when no longer needed.
Security teams must assume that criminals will target the instructions, identity, and agent account rather than only the card number. A manipulated agent operating with valid credentials can create transactions that look technically authorised.
Liability Must Be Clear Before Scale
Agent payments raise difficult questions when a customer rejects a purchase. Did the software exceed its instruction, did a merchant present misleading information, did the bank approve an invalid action, or did the customer grant authority too broadly?
Dispute processes need evidence showing the original request, constraints, product information, agent reasoning, approval path, and final transaction. Customers also need a human route that does not require them to debate an opaque model.
Banks should distinguish recommendation, preparation, and execution in the interface. A customer must know whether an agent is suggesting an item, filling a basket, or committing funds.
The DBS pilot represents a wider shift from AI as productivity software to AI as an operational participant. Payments are an especially important boundary because errors create immediate financial consequences.
Agent-driven commerce will scale only if delegation feels narrower and safer than handing over a card. Tokenisation and issuer approval provide the technical foundation. Trust will come from customers being able to see, limit, understand, and reverse what their agent is allowed to do.
