How AP2 works
When a person buys something online, the shop and the bank can both see evidence that a human was there: card details entered, a password approved, a device confirmed. An agent buying on someone’s behalf removes that evidence, leaving everyone else to take the agent’s word for it. AP2, the Agent Payments Protocol, is an open specification published by Google with a group of payment and commerce companies to put the missing proof back.
It does that with signed authorisations, usually called mandates. One records the buyer’s instruction before anything is bought — what they want, what they are willing to spend, any conditions attached. Another records the exact basket the agent is about to pay for, so the merchant can check that the cart matches what the buyer actually agreed to. Both are cryptographically signed, so afterwards there is a verifiable record of who authorised what, rather than a claim made by software.
The protocol is written to work across payment methods rather than being tied to one, and to sit alongside other agent standards instead of replacing them. An agent can find a shop and assemble a basket through whatever channel it uses, then rely on AP2 where money actually moves.
Why AP2 matters
Everyone in the chain needs an answer that a chat transcript cannot give. The buyer needs to know the agent cannot spend beyond what was agreed. The merchant needs to know the order is genuine before shipping goods. The bank or gateway needs to know who carries the loss when something goes wrong. A signed mandate gives all three the same piece of evidence, which is the condition for agent purchasing becoming ordinary rather than experimental.
For most businesses the practical significance is not implementation. The decision about how agent purchases get authorised will be made by payment providers and platforms, and it will reach your checkout as a setting rather than a project. Knowing the vocabulary now means you can ask the right question when it does.
Where AP2 goes wrong
The most common error is treating it as a feature to switch on. It is a specification, and it only does anything when the buyer’s agent, the merchant’s checkout and the payment provider all speak it. Until they do, an agent that appears to pay on your site is really just filling a form with stored card details, with none of the protections the protocol describes.
The second is over-broad authority. An instruction along the lines of “buy whatever you think I need” is expressible and dangerous. Narrow mandates — a named item, a spending ceiling, a named shop, an expiry — are the whole point; open-ended ones hand the risk straight back to the buyer.
The third is confusing authorisation with judgement. A signed mandate proves the purchase was permitted. It says nothing about whether it was a sensible purchase, and an agent can be perfectly authorised and still buy the wrong thing.
What to do about it
If you sell online, ask your payment gateway and your ecommerce platform what their plan is for agent-initiated purchases, and treat it as a roadmap question rather than something to build. In Nepal, where the gateway and wallet market is smaller and more local than the markets these standards are drafted for, expect support to arrive later than the announcements suggest.
If you are the one deploying agents, the discipline is the same as any other automation. Give an agent the narrowest permission that finishes the job, keep the record of what it was told to do, and keep a person in the loop wherever money moves. Scoped access and a working approval gate protect you today in a way the protocol cannot yet, and the same rule applies to any AI agent placed inside a business process.
