Mastercard has demonstrated a purchase in which an AI agent searched for a product, assessed the website, and completed payment without the user opening an application or entering card details. The transaction took place inside a framework intended to authenticate both the person and the software acting for them.
The India AI Impact Summit demonstration was controlled, not a public launch. Mastercard says wider deployment depends on regulation and ecosystem readiness. Yet the test captures a major change: checkout becomes delegated authority rather than a moment when a human is necessarily present.
Payment networks, banks, merchants, and enterprise finance teams will need to recognise software as a participant with identity, permissions, and accountability.
Consent Must Persist Until the Transaction
Modern payments already use stored credentials, tokenisation, one-click checkout, and fraud monitoring. Agentic commerce adds the ability to search and choose before payment, following rules established earlier by the user.
Those rules could limit amount, merchant, product category, frequency, geography, or substitutions. The payment system must verify both the account owner’s consent and the agent’s authority at the exact moment money moves.
The distinction between recommendation and execution needs to remain visible. A user should know whether software is comparing products, filling a basket, or committing funds. Permissions should be narrow, easy to revoke, and designed to expire.
For corporate procurement, the same principle applies at larger scale. An agent may reorder standard supplies within budget but require approval for a new vendor or changed terms.
Every transaction needs evidence showing the instruction, identity, token, policy checks, product conditions, and final authorisation.
Merchant Systems May Need to Serve Machines
Web shops are designed for human browsing, visual comparison, and interactive checkout. Agents work better with structured product catalogues, current inventory, explicit pricing, delivery options, and machine-readable return policies.
Merchants with inconsistent data or hidden fees may be filtered out before a person sees them. Accurate APIs could become as important as an attractive storefront.
This may alter competition. Agents can compare many options quickly and optimise towards the user’s stated goals. Brand preference may still matter, but it must be represented in the instruction or supported by reliable product information.
Merchants also need controls that distinguish legitimate automated buyers from scraping, fraud, or abusive purchasing. Rate limits, identity signals, and payment-network verification should work together rather than blocking all machine traffic.
Returns and disputes must remain manageable when the customer never visited the site. The merchant should be able to reconstruct what the agent saw and which terms applied.
Security Risk Moves Towards Delegated Authority
A compromised assistant with payment access could make valid-looking purchases at speed. Fraud systems trained on human interaction may mistake automated behaviour for an attack—or fail to notice a malicious agent using legitimate credentials.
Networks will need models of normal agent behaviour, while banks enforce spending limits and step-up approval for unusual actions. Device, agent, merchant, and customer identity signals all become relevant.
Liability must be clear before adoption scales. If an agent misinterprets a request, responsibility may be disputed among user, agent provider, merchant, bank, and payment network. Regulation will need to define what meaningful consent looks like when execution occurs later and automatically.
Mastercard’s demonstration is less about an AI shopping tool than the trust layer beneath one. The network’s strategic position is to authenticate machine-initiated commerce just as it authenticates human payments today.
If this model matures, checkout may disappear from the customer journey. People will define intent and limits while software handles the transaction. Convenience will depend on something less visible but more important: proof that the agent spending the money is the agent the customer authorised, doing only the job it was given.
