The Agentic E-Commerce Checkout Protocol: An APAC Operator's View

Quick Answer: An agentic e-commerce checkout protocol lets AI agents complete purchases on a shopper's behalf. ACP — co-developed by OpenAI and Stripe — defines a product feed, checkout session API and delegated payment tokens. The merchant stays merchant of record and owns tax, inventory and fraud logic.

Most of the coverage of the agentic e-commerce checkout protocol treats it as a traffic story — a new channel where ChatGPT sends you orders. That framing is wrong, and it leads merchants to the wrong work. Agentic checkout is not a channel. It is a change in who holds the customer relationship at the moment of payment, and a change in which party is liable when the order goes wrong. If you optimise for the traffic and ignore the liability, you will onboard fast and discover the cost at month three, when you are reconciling disputes against an order you never saw a customer for.

Related reading: Salesforce Fin Acquisition: AI Agent Strategy for APAC Teams

I have spent the last decade building commerce systems for listed retail groups, catering operators and dealer networks across Greater China and Southeast Asia. The pattern here is familiar: a US-originated standard arrives, the platform vendors ship support quickly, and the parts that break are the parts specific to how APAC actually takes payment and fulfils orders. That is the lens worth applying.

Related reading: Iroh 1.0 Distributed Systems Protocol: An APAC Build Guide

Related reading: Local LLM Models vs Cloud APIs for Coding: APAC Verdict

Three protocols, but only one decision for most merchants

There are three acronyms in circulation and they are not competitors in the way the headlines suggest.

ACP (Agentic Commerce Protocol) was co-developed by OpenAI and Stripe and published as an open specification in late September 2025, releasing under Apache 2.0 alongside Instant Checkout in ChatGPT (Stripe, "Developing an open standard for agentic commerce"). ACP is narrow and concrete: it defines how an AI agent reads your catalogue, creates a checkout session against your systems, and submits a tokenised payment. The merchant remains the merchant of record.

UCP (Universal Commerce Protocol) is the Google-aligned effort, built around agent-to-merchant interaction more broadly — discovery, cart state, and commerce actions exposed to agents such as Gemini, with Shopify among the early named participants. It overlaps with ACP at the checkout step but reaches further up the funnel.

Related reading: Shopify Plus Noibu Ecommerce Analytics in APAC: What It Fixes

AP2 (Agent Payments Protocol) sits underneath both. Google announced it in September 2025 with support from more than 60 payment and commerce organisations, including Mastercard, American Express and Adyen (Google Cloud, AP2 announcement). AP2 is about cryptographically signed mandates — proving that a human actually authorised an agent to spend a specific amount under specific conditions. It answers the question ACP mostly defers: how do you prove intent when no human clicked "Pay"?

MCP (Model Context Protocol), which appears in the same conversations, is not a commerce protocol at all. Anthropic open-sourced it in November 2024 as a general standard for connecting models to tools and data (modelcontextprotocol.io). Plenty of merchants will expose an MCP server for internal agents — inventory lookups, order status — without ever touching ACP.

Related reading: AI Inference Semiconductor Demand in APAC: What the Data Shows

The practical decision for most merchants is not "which protocol do I back". It is "which agent surfaces are worth being present on, and what does each one require of my stack". ACP is the one with live consumer purchase volume today, so it is where the near-term work sits.

What an agentic checkout protocol actually does

Strip away the narrative and an agentic e-commerce checkout protocol is three capabilities exposed over HTTP.

One: a machine-readable product feed. The agent needs structured data on availability, price, variants, shipping options and return policy — not your rendered PDP. ACP expects a feed with per-variant inventory and fulfilment detail.

Two: a checkout session API. The agent creates a session, updates it (address, shipping method, promotion), and the merchant returns recalculated totals including tax and freight. A minimal request looks like this:

1POST /checkout_sessions HTTP/1.1
2Host: api.merchant.example
3Authorization: Bearer {merchant_api_key}
4Content-Type: application/json
5Idempotency-Key: 9f4b2e1a-7c3d-4f10-b8a1-22c0f9d1e4aa
6
7{
8  "items": [{ "id": "sku_HK_8842", "quantity": 2 }],
9  "buyer": { "email": "[email protected]" },
10  "fulfillment_address": {
11    "country": "HK",
12    "city": "Kowloon",
13    "line_1": "Unit 12, 8 Hok Yuen Street"
14  }
15}

The response carries line items, a totals breakdown, available fulfilment options and a session status. Your system — not the agent — owns the arithmetic. That matters: tax, duty, and shipping surcharge logic for cross-border APAC shipments is where naive implementations leak margin.

Three: payment completion with a delegated token. The agent submits a payment credential vaulted by the payment provider rather than a raw card number, and the merchant captures through its existing processor:

1{
2  "payment_data": {
3    "token": "spt_1QxAgenticDelegated",
4    "provider": "stripe",
5    "billing_address": { "country": "HK" }
6  }
7}

That is the whole shape. The complexity is not in the endpoints — it is in everything downstream of a successful capture.

Shopify Plus merchants get this nearly for free, which is the real problem

Shopify was named in the original Instant Checkout rollout alongside Etsy, with Shopify merchant availability following the initial Etsy launch (OpenAI, Instant Checkout announcement, September 2025). For a Shopify Plus merchant, enabling agentic checkout is largely a settings-and-eligibility exercise rather than an engineering project. Adobe Commerce and SHOPLINE merchants, by contrast, are looking at genuine integration work — building the feed, the session endpoints, and the token handling themselves.

The counterintuitive part: the Plus merchants are more at risk, not less. When enablement is a toggle, nobody is forced to think through the operational consequences. We have seen this pattern before with Shop Pay and with marketplace channel apps — the integration is trivial, so it ships without a corresponding change to fraud rules, CS scripting, or returns handling.

The specific things that break when an order arrives from an agent surface:

• Attribution collapses. Your order has no referring session, no UTM chain, no on-site behaviour. If your paid media decisions run off last-click in GA4, a growing slice of revenue becomes invisible to the model that allocates your budget.
• Fraud scoring loses its features. Device fingerprint, session velocity and behavioural signals are the backbone of most rules engines. An agent order presents none of them in a recognisable form. Expect false positives first — legitimate orders blocked — before you see genuine fraud.
• Post-purchase ownership is ambiguous. The customer bought inside ChatGPT. Where do they go when the parcel is late? Your CS team will receive tickets from people who have no order confirmation in their inbox and no account on your store.
• Marketing consent does not transfer. A delegated-payment order does not grant you a marketing relationship. If your LTV model assumes every first order is the start of an email sequence, agentic orders break the assumption.

None of these are reasons not to enable it. They are reasons to enable it with a named owner and a 60-day review.

Why APAC payment rails strain the assumptions in ACP

ACP as shipped assumes a card-centric world. Delegated payment tokens, network tokenisation, 3DS exemptions — all of this maps cleanly onto Visa and Mastercard rails. That assumption holds in the US and reasonably well in Australia and Hong Kong. It holds far less well across the rest of the region.

Consider the composition. Digital wallets accounted for roughly two-thirds of Asia-Pacific e-commerce transaction value in 2024 and are projected to keep gaining share through 2030 (Worldpay Global Payments Report). In Indonesia, Vietnam and the Philippines, the dominant instruments are wallets, bank transfers and QR-code schemes — GoPay, OVO, MoMo, GCash, PromptPay, PayNow, FPS. None of these have a delegated-token story equivalent to a vaulted card today.

The practical consequences:

• Agentic checkout in Southeast Asia will be card-only for a while. That is a self-selecting, higher-income, lower-volume cohort in most of these markets. Your agentic conversion rate will look odd against your overall mix because the payment method mix is not comparable.
• Cash on delivery is incompatible. COD remains meaningful in Vietnam, the Philippines and Indonesia. A protocol built around pre-authorised payment has nothing to say about it.
• Cross-border duty and tax is on you. A buyer in Singapore ordering from a Hong Kong merchant triggers GST obligations under Singapore's low-value goods regime, in force since January 2023 (IRAS). The agent will not compute that. Your session endpoint must, and it must do so in the same call where the agent expects totals within a few hundred milliseconds.
• Regulatory posture is unsettled. The Monetary Authority of Singapore and the Hong Kong Monetary Authority have both published guidance on AI use in financial services, but authorisation-by-agent is not squarely addressed in either framework. AP2-style signed mandates exist precisely because this gap will need closing.

For a Greater China jewellery retailer we worked with, the binding constraint on any new channel was never the integration — it was that high-value SKUs require identity verification and, for certain categories, in-person collection. An agent cannot complete that flow. The honest answer for product lines like that is to expose discovery and reservation, not checkout.

What agentic commerce demands from your product data

This is where the work actually is, and it is unglamorous.

Agents do not browse. They query, filter and compare against structured attributes. If your catalogue says "Colour: Navy / Blue / Midnight" across three product lines, an agent filtering for blue returns a third of your inventory. If your dimensions live in free-text descriptions, you do not appear in "fits a 15-inch laptop" comparisons at all.

Concretely, the minimum viable catalogue hygiene for an agentic e-commerce checkout protocol integration:

• Per-variant availability, not product-level. Agents commit to a specific SKU. Overselling a variant inside an agent-initiated order is a dispute you cannot easily resolve, because there is no customer session to recover.
• Normalised attribute taxonomy. One vocabulary for colour, size, material, compatibility. Map your legacy values once, then enforce at ingestion.
• Explicit shipping and return terms per destination. Agents increasingly surface return windows and delivery estimates as comparison criteria. "Contact us" is a non-answer that drops you from consideration sets.
• Structured data on the storefront too. Schema.org Product and Offer markup remains how non-ACP agents and AI search surfaces read your catalogue. Google's own guidance on product structured data has not changed; it just matters more now.
• Price parity across surfaces. If your agent feed and your PDP disagree, the agent's session creation fails or, worse, succeeds at the wrong price.

A useful test: export your feed, hand it to an engineer who has never seen your catalogue, and ask them to answer five comparison questions a shopper might ask. If they cannot, neither can the agent.

A readiness sequence that does not require betting on a winner

The protocols will consolidate, and nobody knows into what. Build the things that are valuable regardless.

Weeks 1–3: instrument and attribute. Create a distinct order source for agent channels in your data layer and your warehouse. Get it into the same tables your finance team already trusts. Without this you cannot answer whether agentic revenue is incremental or cannibalising.

Weeks 3–6: catalogue normalisation. Attribute taxonomy, per-variant inventory accuracy, destination-level shipping and returns. This work pays off in marketplace feeds, AI search visibility and on-site filtering whether or not you ever enable ACP.

Weeks 6–9: payments and tax logic in the session path. Confirm your tax engine can return totals synchronously for your top destination countries. Decide, explicitly, which markets you will accept agent orders from. Geo-restricting is a legitimate choice.

Weeks 9–12: operations. Fraud rules adjusted for missing device signals. CS runbook for "I bought this in ChatGPT". Returns flow that works without a customer account. A named owner for the channel.

Shopify Plus merchants can compress this considerably, because the session and payment layers are handled. But do the instrumentation and operations work anyway — they are the parts the platform cannot do for you.

The honest assessment: who should skip this

Agentic commerce volume is early. ChatGPT reached roughly 800 million weekly users by October 2025 according to OpenAI's own DevDay figures, which is a large surface — but weekly users and transacting shoppers are very different numbers, and nobody has published credible conversion data on agentic checkout at scale. Anyone telling you the revenue share today is anything other than small is selling something.

So: if you are a sub-US$2M GMV merchant with a catalogue that is not normalised, your return on fixing product data for on-site search and marketplace feeds is higher than your return on agent readiness, and it is the same work. Do that first. If your business depends on COD, wallet payments, or regulated categories requiring identity checks, agentic checkout will not reach most of your customers for several years. If you sell considered high-ticket goods where the sales conversation is the product — bespoke jewellery, dealer-mediated equipment, anything with a fitting or a site survey — exposing checkout to an agent solves a problem you do not have.

The merchants for whom this is urgent are narrower than the discourse suggests: high-SKU-count catalogues in replenishable or comparison-heavy categories, card-dominant markets, with clean product data and a Shopify Plus or Stripe-based stack already in place. If that is you, the integration cost is low and the option value of being present early is real.

For everyone else, the right posture is to treat the agentic e-commerce checkout protocol work as a forcing function for catalogue and operational discipline you should have anyway — and to build the instrumentation now, so that when the volume does arrive, you can tell the difference between a new customer and a cannibalised one. The protocols will change. The quality of your product data will not stop mattering.

If you are weighing whether agentic checkout belongs on your 2026 roadmap, Branch8 runs commerce architecture reviews across Shopify Plus, Adobe Commerce and SHOPLINE for merchants operating across Hong Kong, Singapore, Taiwan, Australia and Southeast Asia — including the unglamorous catalogue and tax-logic work that determines whether any of it functions.

Sources

• Stripe — Developing an open standard for agentic commerce
• Agentic Commerce Protocol — Specification
• OpenAI — Instant Checkout and agentic commerce
• Model Context Protocol — Introduction
• Google Cloud — Agent Payments Protocol (AP2)
• Worldpay — Global Payments Report
• IRAS — GST on imported low-value goods
• Shopify — Developer documentation