Agentic payments · Jun 21, 2026 · 2 min read
Google's AP2 (Agent Payments Protocol), Explained
AP2 is Google's open protocol for AI-agent payments, built on signed 'mandates' that prove what a human authorized. How the mandate chain works, how AP2 pairs with UCP, and what it changes for merchants and disputes.
When Google published AP2 — the Agent Payments Protocol — with dozens of payments partners behind it, the framing was deliberately unglamorous: a trust layer for transactions a human doesn't directly click. That unglamorous layer is where agentic commerce either earns the payment industry's confidence or stalls.
AP2's core idea is worth understanding even if you never read a spec, because it changes what 'proof' means in a dispute.
The mandate chain
AP2 models delegation as verifiable credentials — tamper-evident, signed digital records called mandates:
- Intent mandate — what the human asked for and the bounds they set ('a carry-on under $300, buy when it drops'), signed up front. This is what lets an agent act later without the human present.
- Cart mandate — the specific final cart: items, price, terms, merchant. For human-present flows this is signed at confirmation; for delegated flows the intent mandate authorizes its creation.
- Together they form an auditable chain from 'what was authorized' to 'what was charged' — presentable evidence when anything is contested.
Deliberately agnostic
Two design choices define AP2's ambitions. It's payment-method agnostic: the mandate grammar wraps cards, bank transfers, and stablecoin rails alike — there's a published extension pairing AP2 with x402 for crypto-settled agent payments. And it's agent-framework agnostic: any assistant, browser agent, or shopping bot can present mandates, which is what a web-scale standard requires.
AP2 also slots as the payments companion to the Universal Commerce Protocol: UCP handles how an agent shops a store; AP2 handles how it proves it was allowed to pay. Same coalition logic — platforms and networks agreeing on plumbing so no single assistant owns the toll booth.
What changes for a merchant
Mostly, evidence quality. Today a card-not-present dispute is your word against a chargeback code. A mandate-backed agent order arrives with signed proof of authorization and an approved cart — a better artifact than most human orders carry. The trade is discipline on your side: the cart mandate freezes price and terms at authorization, so stale availability, drifting prices, or ambiguous policies turn into documented discrepancies instead of quiet edits.
Which lands the preparation exactly where every other agent surface already put it: true data, canonical policies, labeled agent orders. If your store is legible enough for an assistant to recommend it honestly, it's legible enough to transact under AP2. That legibility layer — probe-tested, synced, measured against your own baseline — is the product Kinect runs.
Frequently asked questions
Is AP2 competing with OpenAI's ACP?
They sit at different layers more than opposite ends: ACP is a commerce protocol with a built-in payment approach (Stripe's delegated tokens); AP2 is a payments-authorization protocol designed to sit under many commerce flows, UCP first among them. Expect bridges, not a winner-take-all.
When do merchants need to support AP2?
You don't 'support' it directly — PSPs, platforms, and networks do. Your exposure is indirect: better dispute evidence, stricter truth requirements on price/availability, and new order metadata to route into reporting.
Do mandates mean agents can buy without a human?
Only within signed bounds the human set in advance — that's the intent mandate's whole job. Unbounded agent spending isn't a thing any of these protocols permit by design.
Related reading
See Kinect on your store
An AI sales rep on your storefront and an agent-ready store behind it — trained on your catalog and policies, live the same day, measured against your own baseline.