Consumers increasingly ask an AI assistant to find and buy a product instead of browsing shops. Merchants who only optimise for visibility (see REC-011) lose the sale at the moment the agent wants to check out. Agentic commerce enablement exposes catalogue, price, stock, cart and checkout through open agent protocols and accepts delegated payment credentials, so the shop stays merchant of record while the buyer-side agent does the browsing. It is early: volumes are small, protocols move fast.
Problem & value
- Today: AI assistants recommend products but send users to a shop's human checkout. Agents scraping product pages are brittle, bot defences block them, and card data cannot safely be handed to an agent. Shops have no identity, consent or liability model for agent-placed orders.
- AI value: a clean machine interface (feed + checkout API + tokenised payment) lets an agent complete the purchase within the conversation, with the shop keeping pricing, tax, fulfilment and customer relationship. New demand channel without paying a marketplace to be the seller of record.
- Evidence level: real deployments exist but are young. ACP is maintained by OpenAI and Stripe (Apache 2.0, spec in beta, stable version 2026-04-17) [1]; UCP is an open standard from Google with Shopify, Etsy, Wayfair, Target, Walmart and others, and keeps the business as merchant of record [2]. Several secondary sources report that ChatGPT Instant Checkout was retired in March 2026 (exact date varies by source) while the protocol continued; check current status before planning around a single surface. No reliable conversion or revenue uplift figures are public; measure your own.
Variants
- Protocol-native merchant: implement ACP and/or UCP endpoints directly (large retailers, marketplaces).
- Platform-enabled SMB: switch on the feature in Shopify, commerce platforms or the PSP; no custom code.
- PSP-side view: acquirer or PSP issues and accepts delegated tokens (Stripe Shared Payment Tokens, network agentic tokens), runs agent authentication and agent-aware fraud rules.
- Card-network track: Visa Intelligent Commerce / Trusted Agent Protocol and Mastercard Agent Pay add network-level agent identity and agentic tokens, independent of the shopping surface.
- B2B and digital goods: AP2-style mandates for repeat or agent-to-agent purchasing of services and subscriptions.
- Out of scope: the consumer's own agent (PERS-011), the shop's own chatbot (CONV-006), pure visibility (REC-011).
How it works
- AI assistant (buyer agent)
- discovers merchant (feed / UCP profile / MCP-style endpoint)
- product search, price & stock query
- cart creation
- shipping & tax quote
- buyer confirms in the assistant
- delegated payment token (SPT / network token / AP2 mandate)
- merchant checkout API creates order, PSP authorises (SCA by issuer where required)
- fulfilment, e-mail receipt
- returns / support via order ID
- Catalogue layer: complete, structured product data (GTIN, variants, availability, delivery time, total price incl. VAT and shipping). Quality here decides whether an agent can rank and buy the product.
- Checkout layer: ACP exposes Agentic Checkout and Delegate Payment APIs [1]; UCP organises commerce into capabilities (discovery, checkout, order management) with extensions such as discounts [2]; Google AP2 adds signed mandates proving what the user authorised [3].
- Payment layer: the agent never sees the card number; it passes a scoped, single-use token tied to merchant and amount. Visa and Mastercard add agent identity and cardholder authentication (passkeys) on the issuer side [4][5].
- AI components: mostly protocol engineering. LLMs help with product-data enrichment, agent-traffic classification and answering post-purchase queries. Human-in-the-loop: the buyer confirms purchase in the agent; merchant staff handle disputes, refund exceptions and fraud reviews.
Data preparation
The core data is the product catalogue: GTINs, titles, attributes, prices, stock, delivery times, and the return, shipping and consumer-law policies. In an SMB it typically lives in the shop system, the ERP or merchandise system, supplier Excel and CSV files, and PDFs from manufacturers. Typical gaps are missing GTINs, inconsistent attributes and units, thin or duplicated descriptions, stale stock and price data, images without alt text, and policies that exist only as scattered documents. Agents compare offers on structured data, so incomplete attributes mean lost visibility.
Concrete steps: export the catalogue from the shop and ERP (PREP-005), put it into tidy tables (PREP-011), and define attribute names and units in a glossary (PREP-012). Add tags and metadata (PREP-010), fill gaps from supplier data, and validate feeds against the protocol schema. Write return and delivery policies as clear, authoritative texts (PREP-013) so that agents and support answer consistently. Keep stock and price synchronisation near real time. Expect roughly a third of the effort to go into catalogue quality, and ongoing maintenance afterwards. For a readiness check see GUIDE-003.
Implementation paths
Which path for whom
- SMB: A. Shop platforms (Shopify, commercial shop systems with ACP/UCP connectors) already offer protocol support; the process is standard and a hosted product is cheapest.
- Enterprise: B. Own commerce stacks and PIM/OMS landscapes need custom feed, checkout and risk integration on APIs, and agent traffic is strategic.
- Developer: B. Implementing the protocol endpoints and token handling on cloud APIs and payment-provider SDKs is the natural route for agencies and platform builders.
Path A — Buy / enable on platform (recommended for SMB)
- Commerce platforms: Shopify (participant in both ACP and UCP rollouts), other SaaS shops with agent-checkout plugins; check which agent surfaces are live in your country.
- PSPs: Stripe (Shared Payment Tokens, Agentic Commerce Suite), Adyen (EU), Checkout.com, Worldline (EU), Nuvei, with network agentic tokens from Visa and Mastercard. Verify EU availability and SCA handling.
- Feed/catalogue tools: product-information systems (Akeneo, EU) and feed managers for clean attributes.
- Choose A when: you run a standard shop and want to test the channel without engineering.
Path B — Build on APIs
- Implement ACP checkout and delegate-payment endpoints or UCP capabilities on your own backend, using your PSP's token acceptance. Reuse existing cart, tax and order services behind a thin protocol adapter.
- Agent authentication: verify agent signatures (e.g. Visa Trusted Agent Protocol, HTTP message signatures), rate-limit, and log the agent identity per order.
- LLM APIs (EU-region Mistral, Azure OpenAI EU, Claude, Gemini) for catalogue enrichment and support replies.
- Choose B when: custom ERP/stock logic, marketplaces, or several brands behind one checkout.
Path C — Self-host
- Protocols are open source (ACP and UCP specs on GitHub), so a self-hosted adapter in front of Medusa, Saleor (both open source) or a custom shop is feasible. Payment acceptance still requires a licensed PSP/acquirer.
- Small open-weight models (Mistral, Qwen, Llama class) for enrichment and agent-traffic triage on HW1–HW2; not required for the protocol itself.
Path D — Custom training
Rarely. Possible only for agent-vs-human traffic classifiers or fraud models trained on your own order and session history once agent volume is meaningful.
State of the art (2026)
- Protocol landscape: ACP (OpenAI/Stripe) [1], UCP (Google and retail partners) [2], AP2 for payment authorisation [3], plus network schemes: Visa Intelligent Commerce / Agentic Ready and Mastercard Agent Pay [4][5]. They are largely complementary (UCP for the journey, AP2 and network tokens for payment), but no single winner exists.
- Europe: Visa launched its Agentic Ready programme in March 2026, starting in Europe with more than 20 issuing partners [4]; Mastercard and ING reported first European production agentic payments [5]; Revolut and Visa reported France's first passkey-authenticated agentic card payment in a live checkout in September 2026, as a controlled test not yet enabled in the app [9]. Most are pilots, not consumer-scale launches.
- Limits: spec churn (ACP is beta), few agent surfaces with real traffic, unclear liability allocation for wrong orders, weak product data causing wrong variants, prompt-injection of buyer agents by hostile page content, and merchants losing the direct customer relationship and upsell surface.
Shared resources & combinations
- REC-011 (be found by agents) is the demand side; this use case is the transaction side.
- ANOM-001 payment fraud models need agent-aware features; AGENT-020 e-commerce operations shares catalogue and reply automation; CONV-006 reuses the same product data; AGENT-003 covers agents driving UIs without protocols (the fallback this use case replaces); AGENT-025 for monitoring your own agents; PERS-011 is the buyer's counterpart.
- Shared building blocks: product data model, PSP tokenisation, webhook/order workflow, bot-management.
Effort, cost & KPIs
- Path A: days to weeks to enable; cost is PSP/platform fees and feed clean-up.
- Path B: 2–4 months for protocol adapter, agent authentication, test with each agent surface, plus ongoing spec maintenance.
- Path C: similar to B plus PSP integration; hosting is minor.
- KPIs: share of orders from agents, feed completeness and rejection rate, checkout success rate, authorisation rate (incl. SCA challenges), chargeback and fraud rate for agent orders vs web, return rate, margin after agent-channel fees.
People & change
Involve management for the strategic decision and the margin and brand questions, the sales or e-commerce lead as process owner, and IT and information security for endpoints, payment tokens and bot or fraud controls. The data protection officer clarifies roles with the platform and payment provider (AVV, Art. 28 GDPR), what customer data reaches third-party agents, and whether a DPIA is needed. Legal and compliance review consumer-law duties, merchant-of-record liability and the contracts with agent platforms; suppliers must deliver clean product data.
Works council: no in the typical case, because the use case concerns external customers and agents and does not monitor employee behaviour (§87(1) no. 6 BetrVG). It becomes relevant only if the support or fraud tooling evaluates individual employees; still inform the council in good time under §90(1) no. 3 if work processes change.
Start with a pilot on a small product range or one agent channel, with spending limits and human review of unusual orders. Typical concerns are loss of customer relationship, margin pressure and fraud; address them through clear brand and price rules and monitoring. Train shop and support staff on agent orders (AI Act Art. 4 requires measures supporting AI literacy). See GUIDE-009, GUIDE-014 and GUIDE-010.
Risks, regulation & compliance
- EU AI Act: minimal risk for the merchant's protocol layer; AI literacy (Art. 4) applies. Chatbot-style replies to consumers carry Art. 50 transparency duties. Hidden manipulation of agents or users could touch Art. 5 (manipulative techniques), so do not craft deceptive agent-directed content.
- PSD2/SCA: the issuer remains responsible for authenticating the cardholder; delegated agent payments rely on tokens plus issuer-side authentication (passkeys, 3-D Secure) and agent identity. Check exemptions with your PSP and the evolving PSD3/PSR text rather than assuming that agent payments are exempt.
- Consumer law: the shop is the trader: price transparency, withdrawal right (14 days), pre-contractual information, Consumer Rights Directive and Digital Services Act duties apply even if the information is shown in a third-party agent.
- GDPR: orders carry names, addresses and purchase histories; clarify controller roles with the agent provider, limit data sent to agents, and document processing agreements with PSP and platforms.
- Security/fraud: agent impersonation, token replay, price manipulation, prompt injection via product text; mitigate with agent signature verification, token scoping, price re-validation at checkout and order limits for new agents.
Getting started
- Audit product data (GTIN, attributes, total price, delivery times, return policy) and fix the gaps.
- Ask your platform and PSP which agent protocols and surfaces they support in the EU today.
- Pilot with a limited assortment and cap on order value; log agent identity per order.
- Add agent-traffic rules to bot management and fraud screening instead of blocking all bots.
- Review terms, return process and consumer-law texts with counsel for agent-placed orders.
Sources
- Agentic Commerce Protocol (OpenAI/Stripe) repository — https://github.com/agentic-commerce-protocol/agentic-commerce-protocol
- Google Developers Blog — Universal Commerce Protocol — https://developers.googleblog.com/under-the-hood-universal-commerce-protocol-ucp/
- Google AP2 (Agent Payments Protocol) — https://github.com/google-agentic-commerce/AP2
- Visa Intelligent Commerce — https://developer.visa.com/capabilities/visa-intelligent-commerce
- Mastercard Agent Pay — https://www.mastercard.com/us/en/business/artificial-intelligence/mastercard-agent-pay.html
- Stripe — Agentic commerce / Shared Payment Tokens — https://docs.stripe.com/agentic-commerce
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Directive (EU) 2015/2366 (PSD2) — https://eur-lex.europa.eu/eli/dir/2015/2366/oj
- The Paypers — Revolut and Visa complete France's first agentic card payment (Sept 2026) — https://thepaypers.com/payments/news/revolut-and-visa-complete-frances-first-agentic-card-payment