Know Your Agent Payments: What Businesses Must Record

What the three payment firms announced

Reuters reported on 10 September that Ant International, Mastercard and Visa had launched an initiative to develop common standards for identifying and verifying AI agents that make purchases for users. Their proposed Know-Your-Agent framework is intended to let card networks, digital wallets, agent platforms and online marketplaces recognise trusted agents across payment ecosystems without replacing each participant's approval and risk-management process.

The initiative builds on Visa's Trusted Agent Protocol, Mastercard Verifiable Intent and Ant International's Agentic Mobile Protocol. The companies are collaborating through BuildFin.ai, a platform convened by the Monetary Authority of Singapore. The announcement did not publish a final specification, implementation timetable or liability model, so it should be treated as standards work in progress rather than a production guarantee.

Agent identity does not prove transaction authority

A merchant may need to know that a request came from a legitimate agent, but that does not establish what the user authorised. The agent could be genuine while the instruction is stale, exceeds a spending limit, targets the wrong merchant or reflects a compromised account. Identity answers who acted. A mandate must answer for whom, for what, within which limits and until when.

Google's open Agent Payments Protocol makes the distinction concrete. It describes cryptographically signed mandates that preserve a user's initial intent and, where a person approves a specific cart, the exact items and price. Different standards may implement this differently, but the operating principle is useful now: bind identity, intent and transaction state instead of treating a valid agent credential as open-ended permission.

Use a six-part transaction evidence record

Before allowing an AI agent to buy, renew, book or change a paid service, require one record with six linked elements:

• Principal: identify the person or organisation delegating authority and the account or role from which it came.
• Agent: record the agent identity, provider, version, technical credential and environment that executed the action.
• Mandate: capture the approved objective, merchant or category, amount, time window, prohibited actions and any need for fresh human approval.
• Decision: retain the material options considered, policy checks, approval event and reason the selected transaction was permitted.
• Settlement: bind the mandate to the final merchant, item, amount, currency, payment token, receipt and resulting business state.
• Control: define expiry, revocation, exception handling, dispute ownership and the recovery path if the purchase is wrong or incomplete.

Test disputes before increasing autonomy

An agent-payment pilot should not be judged only by checkout success. Run deliberate failures: a price changes after approval, an item is substituted, a credential is revoked mid-flow, the same mandate is replayed, or a service is purchased outside the allowed region. Confirm that the transaction stops safely and that investigators can reconstruct the authority chain without relying on model reasoning or a vendor dashboard that may later change.

Start with narrow categories, low limits and reversible purchases. Expand authority only when accepted transactions, policy breaches, manual interventions, disputes and recovery time are measured against a defined exposure.

What business leaders should do next

Ask finance, procurement, security and product teams to approve one shared evidence schema before connecting an agent to a payment method. Make the mandate machine-readable, keep a human-readable summary, and require every connector or payment provider to return enough data to reconcile the instruction with the settled outcome.

ELYMENT AI's analyses of Meta Muse, AWS Agent Registry and GPT-6 Astra extend the same discipline across workplace boundaries, agent ownership and observable actions. ELYMENT AI can help turn the six-part record into a practical approval policy, transaction log and test plan for a specific business workflow.

Sources

• Reuters: Visa, Mastercard and Ant International launch agent trust initiative (10 September 2026) - Independent reporting on the Know-Your-Agent initiative, participating firms, existing protocols, intended interoperability and BuildFin.ai collaboration.
• Google Cloud: Agent Payments Protocol and verifiable mandates (16 September 2025) - Primary technical explanation of agent-payment authorisation, authenticity, accountability, signed intent mandates and cart mandates.
• Axios: Visa Trusted Agent Protocol for agentic commerce (14 October 2025) - Independent reporting on Visa's protocol for distinguishing legitimate shopping agents from malicious bots and its intended use beyond Visa's ecosystem.

Continue learning

• Meta Muse: control the boundary between personal agents and workplace systems
• AWS Agent Registry: connect agent identity to ownership and authority
• GPT-6 Astra: audit observable actions rather than model reasoning

Frequently asked questions