Branch8

Agentic Commerce Protocol ChatGPT Checkout: What APAC Retail Must Do

Matt Li
November 16, 2026
10 mins read
Agentic Commerce Protocol ChatGPT Checkout: What APAC Retail Must Do - Hero Image

Key Takeaways

  • ACP is an open spec from OpenAI and Stripe — not an OpenAI-only pipe.
  • Agents require a single API call returning tax- and shipping-inclusive totals.
  • Shared Payment Tokens shift fraud and chargeback assumptions you don't control.
  • APAC wallet rails lack a delegation model; build the API layer protocol-agnostic.
  • Your Google Merchant Center feed defects will suppress AI discovery identically.

Quick Answer: The Agentic Commerce Protocol is an open standard from OpenAI and Stripe that lets AI agents create checkout sessions and submit delegated payment tokens directly to merchants. ChatGPT Instant Checkout runs on it. Merchants stay merchant of record, handling capture, tax, fulfilment and support.


A listed Greater China jewellery retailer asked my team a question last quarter that would have sounded absurd two years ago: "If a customer asks ChatGPT to buy a pendant under HK$5,000, do we appear — and can they pay without leaving the chat?"

Related reading: Customer Data Platform CDP 2026 Trends: An APAC Build Guide

The honest answer at the time was no on both counts. Their product feed was tuned for Google Shopping, their checkout was a bespoke flow behind three layers of promotion logic, and nobody in the organisation owned "AI channel" as a line item. That gap is what the agentic commerce protocol ChatGPT checkout stack now exposes. OpenAI shipped Instant Checkout in late September 2025 and open-sourced the Agentic Commerce Protocol (ACP) it co-developed with Stripe, which means buying inside a chat window stopped being a demo and became an integration decision with a spec, endpoints, and a payment token format.

Related reading: Salesforce Marketing Cloud Agentic CDP: An APAC Reality Check

For retailers across Hong Kong, Singapore, Taipei, Sydney and Auckland, this is less about chasing a shiny channel and more about whether your commerce backend can be driven by something that isn't a human with a browser.

What OpenAI actually shipped, stripped of the hype

Instant Checkout launched on 29 September 2025, initially for US ChatGPT users buying from Etsy sellers, with over a million Shopify merchants named as the next wave, according to OpenAI's own announcement. The mechanics matter more than the launch partners:

Related reading: Shopify Q1 Earnings AI Revenue Impact: An APAC Merchant View

Related reading: Shopify Plus Analytics Monitoring in APAC: Beyond Dashboards

  • The buyer discovers a product inside ChatGPT, taps buy, and confirms the order in the conversation.
  • The merchant remains the merchant of record. You handle payment processing, tax, fulfilment, returns, and customer service.
  • OpenAI takes a commission on completed purchases and has stated that product placement in results is not paid for — ranking is organic to the model's relevance judgement.
  • The underlying protocol, ACP, is published openly so any agent and any merchant can implement it, not just OpenAI and Stripe.

That last point is the strategically important one. Stripe and OpenAI released ACP as an open standard, and Stripe's documentation frames it as payment-method and processor agnostic. It is not an OpenAI-only pipe.

The scale argument is straightforward. OpenAI reported roughly 800 million weekly ChatGPT users at DevDay in October 2025. Adobe Analytics found that traffic to US retail sites from generative AI sources grew 1,200% between July 2024 and February 2025, and that AI-referred visitors converted at a lower rate but browsed more pages per visit — early-funnel behaviour that checkout inside the assistant is designed to close.

The protocol has two halves, and most teams only think about one

ACP splits into an Agentic Checkout Spec and a Delegated Payments Spec. Merchants tend to fixate on the first and get blindsided by the second.

Agentic Checkout: your store becomes an API

You expose a small set of REST endpoints that an agent calls on the buyer's behalf — create a checkout session, update it, complete it, cancel it. The agent never scrapes your storefront. It negotiates with your commerce engine directly.

1POST /checkout_sessions HTTP/1.1
2Host: api.yourstore.com
3Authorization: Bearer <agent_api_key>
4Content-Type: application/json
5Idempotency-Key: 9c1f0f2a-7b2e-4c0a-9d3f-5f0e2a1b8c44
6
7{
8 "items": [{ "id": "SKU-PEND-0042", "quantity": 1 }],
9 "buyer": { "email": "[email protected]" },
10 "fulfillment_address": {
11 "name": "Chan Siu Ming",
12 "line_one": "12 Queen's Road Central",
13 "city": "Hong Kong",
14 "country": "HK",
15 "postal_code": "000000"
16 }
17}

The response must return the full economic truth of that basket — line totals, tax, shipping options, discounts applied, and any blocking messages:

1{
2 "id": "cs_01HZX9",
3 "status": "ready_for_payment",
4 "currency": "hkd",
5 "line_items": [{
6 "id": "SKU-PEND-0042",
7 "base_amount": 480000,
8 "discount": 48000,
9 "subtotal": 432000,
10 "tax": 0,
11 "total": 432000
12 }],
13 "fulfillment_options": [{
14 "type": "shipping",
15 "id": "sf_express_next_day",
16 "title": "SF Express next day",
17 "subtotal": 3000,
18 "total": 3000
19 }],
20 "totals": [{ "type": "total", "amount": 435000 }],
21 "messages": [],
22 "links": [{ "type": "terms_of_use", "url": "https://yourstore.com/terms" }]
23}

If you cannot compute a landed total — including a tiered member discount, a bundle rule, and a courier rate for a Kowloon address — in a single synchronous API call, you cannot service agentic checkout. That is the real integration test, and in my experience with multi-brand retail groups across Greater China, it is where legacy promotion engines fail. The discount logic lives in the storefront JavaScript or a POS middleware layer, not in anything an external caller can query.

Delegated Payments: a token, not a card

The buyer's card stays with the agent's payment provider. What reaches you is a Shared Payment Token — a single-use, scope-limited credential with an amount cap and an expiry, which you then charge through your own processor. Stripe's implementation lets merchants accept these tokens as a payment method on an existing Stripe account; non-Stripe processors can implement the spec directly.

1curl https://api.stripe.com/v1/payment_intents \
2 -u "sk_live_...:" \
3 -d amount=435000 \
4 -d currency=hkd \
5 -d confirm=true \
6 -d "payment_method_data[type]=card" \
7 -d "payment_method_data[card][token]=spt_1Oabc..."

The trade-off is explicit: you gain a buyer who never saw your site, and you accept a transaction where you did not control the device fingerprint, the session history, or the 3DS challenge. Liability and chargeback treatment for agent-initiated payments is still settling. Visa launched Intelligent Commerce in April 2025 and Mastercard announced Agent Pay the same week, both specifically to issue agent-scoped payment credentials with verifiable consent trails — a signal that the networks consider the current authentication model incomplete.

Ready to Transform Your Ecommerce Operations?

Branch8 specializes in ecommerce platform implementation and AI-powered automation solutions. Contact us today to discuss your ecommerce automation strategy.

Why this is an operations problem before it is a channel problem

The instinct in most boardrooms is to treat this as marketing: get listed, get traffic, count orders. That underestimates what breaks.

An agent calling your checkout API behaves nothing like a shopper. It will create sessions it abandons. It will retry on timeout — which is why idempotency keys are mandatory in the spec, not optional. It expects inventory truth within seconds, not the fifteen-minute cache your Google Merchant Center feed tolerates. And it will surface any inconsistency between your feed price and your API price directly to the buyer as a failure.

I worked through this pattern with a Hong Kong catering group running a Shopify Plus storefront alongside an on-premise POS. Their product feed was accurate for in-store SKUs and stale for online-only bundles. A human shopper never noticed — the storefront re-queried at add-to-cart. An agent would have quoted a price, generated a payment token at that amount, and failed at capture. The fix was not AI-related at all: it was consolidating price and availability into a single source of truth exposed over an API. Which is the same work you would do for a marketplace integration, a dealer portal, or a WhatsApp commerce flow.

That is the useful frame for APAC operators. Agentic checkout is the forcing function for headless-grade data hygiene you probably already needed.

Related reading: Hetzner Cloud Pricing vs APAC Infrastructure Cost Reality

How does this play out differently across Asia-Pacific?

Instant Checkout launched US-first, and the gating factor for the region is not technical readiness — it is payments.

In Hong Kong and Singapore, card-on-file through Stripe or Adyen covers the majority of online transactions, so the Shared Payment Token model maps cleanly. Australia and New Zealand are similar, with the added wrinkle that the Reserve Bank of Australia's review of retail payments regulation has been actively examining surcharging and least-cost routing, which affects how you price an agent-initiated card transaction.

Greater China, Taiwan, and most of Southeast Asia are a different story. If 60–70% of your volume arrives via a wallet rather than a card — Alipay, WeChat Pay, GrabPay, GCash, DANA, TrueMoney — the delegated payment token concept has no native analogue yet. ACP is payment-method agnostic by design, but "agnostic" means someone still has to build the wallet-side delegation. Worth noting: Ant International and the regional wallet operators have been shipping their own agent-facing initiatives, and Google's Agent Payments Protocol (AP2), announced in September 2025 with more than 60 payment and commerce partners including Mastercard, PayPal and Adyen, explicitly includes support for stablecoin and account-to-account rails alongside cards.

So there are at least two protocols in play, and betting the implementation on one is premature. The defensible move is to build the commerce API layer once — session creation, authoritative pricing, inventory, fulfilment quoting — and treat ACP, AP2, or a marketplace adapter as thin translation layers on top.

Ready to Transform Your Ecommerce Operations?

Branch8 specializes in ecommerce platform implementation and AI-powered automation solutions. Contact us today to discuss your ecommerce automation strategy.

The feed is still the gate

No protocol helps if the model does not know your products exist. Discovery in ChatGPT runs off structured product data, and OpenAI's merchant documentation specifies a product feed with required attributes — identifiers, price, availability, condition, images, and a clear link — plus recommended enrichment like variant relationships and return policy.

Practical sequence for a mid-sized APAC retailer:

  1. Audit your existing Google Merchant Center feed. If it is clean, you are most of the way there. If it has been limping on 12% disapproval for two years, fix that first — the same defects will suppress AI discovery.
  2. Expose GTINs and MPNs properly. Agents disambiguate products across merchants using identifiers. A jewellery or furniture retailer selling unique SKUs without GTINs needs a consistent internal identifier strategy and rich attributes to compensate.
  3. Write descriptions for retrieval, not for banner copy. The model is matching intent language — "hypoallergenic earrings for sensitive ears under HK$2,000" — against your text. Specification detail outperforms brand adjectives.
  4. Localise currency and availability per market. One feed covering HK, SG and AU with a single currency will produce wrong quotes for two of three buyers.

If you are on Shopify, much of this is handled by the platform's Catalog and the Instant Checkout integration path. If you are on a custom stack or an older SAP Commerce or Magento build, it is genuine engineering work.

What you should not do yet

I would push back on three things I hear regularly.

Do not rebuild your storefront for agents. Agentic traffic is early. Adobe's data shows growth off a very small base, and even bullish projections put AI-mediated commerce at a single-digit share of e-commerce for the next few years. Build the API layer; do not deprecate the website.

Do not accept unbounded agent access. Rate-limit agent API keys, cap token amounts, and log every session with the originating agent identity. An agent that can create 10,000 checkout sessions can also be used to probe your inventory and pricing for competitive scraping.

Do not assume attribution will work. You lose the referrer, the UTM chain, and often the ad-platform conversion signal. Salesforce reported that AI-influenced traffic during the 2024 holiday period contributed a meaningful share of orders while being systematically under-credited in last-click models. Decide now how you will measure the channel, because retrofitting attribution after the orders start arriving is painful.

Ready to Transform Your Ecommerce Operations?

Branch8 specializes in ecommerce platform implementation and AI-powered automation solutions. Contact us today to discuss your ecommerce automation strategy.

A decision checklist you can run this month

Work through these in order. If you stall at any step, that is your actual project.

  • Can your system return a final, tax- and shipping-inclusive total for an arbitrary basket and address via a single API call? If not, the promotion and pricing engine is the first fix.
  • Is there one authoritative source for price and inventory, queryable in under 500ms? Agents time out. Caches lie.
  • Does your product feed pass Google Merchant Center validation with under 2% disapproval? Same data, same defects.
  • Does your processor support delegated or agent-initiated payment tokens, and do you know the chargeback treatment in each market you sell in? Ask your acquirer in writing.
  • Have you assigned ownership? Agentic commerce sits between e-commerce, IT and finance. Unowned, it stalls.
  • Can you identify agent-originated orders in your BI and your CS tooling? A buyer who purchased inside ChatGPT will still email you about delivery.
  • Have you written the policy for what agents may and may not access? Pricing, inventory depth, customer data, loyalty balances — decide before the first integration, not after.

The retailers who benefit from agentic commerce protocol ChatGPT checkout in 2026 will not be the ones who integrated first. They will be the ones whose commerce data was already clean enough that integration took a sprint rather than a replatform — and who treated the protocol as one more consumer of a well-built API rather than a channel requiring its own stack. The wallet-based economies of Southeast Asia and Greater China will get their version of this later than the US, which is, for once, an advantage: you have time to fix the pricing engine before the traffic arrives.

If you want a straight assessment of whether your current commerce stack can serve an agent-initiated checkout — and what it would take if it can't — talk to the Branch8 team.

Sources

FAQ

The Agentic Commerce Protocol (ACP) is an open standard co-developed by OpenAI and Stripe and published in September 2025. It defines how an AI agent communicates with a merchant's commerce system — an Agentic Checkout Spec covering session creation, updates and completion, and a Delegated Payments Spec covering single-use, scope-limited payment tokens. It is processor and payment-method agnostic, so any agent or merchant can implement it.

About the Author

Matt Li

Co-Founder & CEO, Branch8 & Second Talent

Matt Li is Co-Founder and CEO of Branch8, a Y Combinator-backed (S15) Adobe Solution Partner and e-commerce consultancy headquartered in Hong Kong, and Co-Founder of Second Talent, a global tech hiring platform ranked #1 in Global Hiring on G2. With 12 years of experience in e-commerce strategy, platform implementation, and digital operations, he has led delivery of Adobe Commerce Cloud projects for enterprise clients including Chow Sang Sang, HomePlus (HKBN), Maxim's, Hong Kong International Airport, Hotai/Toyota, and Evisu. Prior to founding Branch8, Matt served as Vice President of Mid-Market Enterprises at HSBC. He serves as Vice Chairman of the Hong Kong E-Commerce Business Association (HKEBA). A self-taught software engineer, Matt graduated from the University of Toronto with a Bachelor of Commerce in Finance and Economics.