Knowing who initiated a payment is essential. Knowing whether that actor was authorised to make this particular purchase is a separate problem. AI agents make both questions more visible because software can potentially carry a transaction from a broad instruction to checkout.

What happened

On 10 September 2026, Ant International, Mastercard and Visa announced collaboration on Know Your Agent interoperability for agentic commerce. The work concerns identifying and verifying agents across payment networks while retaining each network’s own verification and risk processes. This was an initiative to develop interoperability, not evidence that a single universal standard was already operating everywhere. Ant’s announcement describes the collaboration.

For companies considering AI-assisted purchasing, this is a useful development to follow. It does not remove the need to define the limits of a purchasing instruction.

Why it matters

Consider an agent asked to replace office equipment. It may correctly identify itself and still choose an unsuitable specification, exceed a budget or accept a delivery date that defeats the purpose of the purchase.

A business therefore needs a mandate alongside an identity. The mandate should describe the permitted category, spending ceiling, relevant supplier conditions and the circumstances that require another decision. It should also expire when the task ends.

Changes deserve particular attention. The original instruction might permit a particular item, while the checkout process offers a substitute, an extended service package or recurring billing. The company should decide in advance whether the agent may accept those changes or must return with a revised proposal.

These are proposed procurement controls. Their detail will depend on the transaction and the organisation’s existing processes. They should not be replaced by a general instruction to obtain the best deal, because price alone may not describe the business requirement.

The bigger shift

My interpretation is that agentic commerce requires a clearer record of delegated intent. Payment infrastructure can help establish the identity of an actor. The business still needs evidence of what that actor was asked to do.

A useful transaction record would connect the request, the approved scope, any amendments and the final action. A reviewer should be able to reconstruct that sequence without relying on an agent’s later summary of its own behaviour.

Aftercare matters too. If an item is returned, a booking changes or a charge is disputed, the organisation needs an owner who can understand and resolve the case. Automating purchase initiation while leaving the exception process undefined can simply move administrative work to another team.

I would include finance, procurement and customer support in a trial where their responsibilities are affected. Their input can reveal conditions that a technically successful checkout misses, such as reconciliation or the information needed to answer a supplier query.

The purpose is consistent with practical AI governance: make authority observable, bounded and reviewable.

My take

Agent identification is a useful foundation. It should make the next layer of questions easier to answer, rather than encourage companies to skip them.

Choose a small class of low-impact purchases and write down the mandate in language a purchaser and an approver both understand. Test a routine order, a substitution and a cancellation before increasing access.

The result should be a complete operating process, including who handles the unfinished transaction. A recognisable agent is a start. A business also needs a defensible account of why that agent was allowed to spend its money.