A brand team preparing a new commerce channel now faces a deceptively simple question: should an AI shopping agent be allowed to buy the product, or should it only help the customer choose?
That is not one permission. Product discovery, recommendation, checkout execution, and post-purchase service expose different data, create different liabilities, and require different owners. The safest way to enter agentic commerce is to treat authority as a ladder—not an on/off switch.
Discovery and transaction are already separate merchant decisions
Current platform designs support that distinction. Shopify lets merchants control catalog access by channel and separately decide whether direct checkout is enabled; when direct checkout is off, shoppers can still be sent to the merchant’s online store. Google’s Universal Commerce Protocol covers the journey from discovery through post-purchase support, but its checkout capabilities are being introduced in phases and the merchant remains the seller of record. Stripe likewise describes agent-ready checkout as a model in which sellers retain control over pricing, inventory, fulfillment policy, tax, refunds, and the customer relationship.
The identity and payment layer is developing too. On August 4, 2026, Cloudflare announced agent identities and virtual wallets with controls such as spending caps, approved-merchant lists, and per-transaction limits. The company also said full wallet access would arrive in the following months. This is useful evidence that infrastructure is being built; it is not proof that every channel is production-ready.
Retail strategies are diverging at the same time. A Wall Street Journal report published on October 2, 2026 described retailers that enable agentic purchases, retailers that allow browsing but not buying, and companies that restrict agent access. That establishes a live strategic split—not an industry-wide transition.
The ONCE Agent Commerce Access Ladder
Move up only when the previous level has reliable data, a named owner, and measurable failure handling.
| Level | Agent authority | Entry requirements | What to measure |
|---|---|---|---|
| 1. Read | Access products, variants, imagery, price, availability, shipping, and returns | Stable SKU and variant IDs; structured fields; authoritative price and inventory; separated B2B and DTC access | Freshness, missing fields, wrong price or stock events |
| 2. Recommend | Compare, filter, and explain why a product fits a request | Approved use cases, limitations, compatibility, and substantiated claims; human review for high-risk claims | Unsupported claims, poor-fit recommendations, qualified handoffs, human corrections |
| 3. Act | Create a cart and initiate or complete checkout | Verifiable agent identity and buyer authorization; amount-, merchant-, and time-limited payment credentials; fresh price, stock, shipping and tax checks; fraud and cancellation controls | Authorization failures, duplicate orders, completion, margin and return-adjusted revenue |
| 4. Resolve | Track orders, request returns, reorder, or escalate support | Least-privilege order access; versioned policies; action logs; clear human escalation for exceptions and disputes | Incorrect refunds, handling time, escalation rate, disputes and customer satisfaction |
The objective is not to reach Level 4 as quickly as possible. A premium, customized, regulated, or high-return product may be better served by strong recommendation followed by merchant-controlled checkout.
Seven gates before an agent can place an order
If any of these questions lacks a clear answer, keep the experience at recommendation plus merchant checkout.
- Product truth: Do names, variants, dimensions, compatibility, and limitations come from one controlled source?
- Price and inventory: Are they recalculated at checkout rather than copied from an earlier answer?
- Identity and consent: Can the merchant verify the agent, the represented buyer, and the amount authorized?
- Payment and fraud: Are credentials restricted by merchant, amount, and time, and can legitimate agents be distinguished from abusive bots?
- Fulfillment and refunds: Who is merchant of record, and who owns tax, shipping, returns, chargebacks, and customer support?
- Attribution: Can recommendation, referral, cart creation, authorization, and purchase be logged separately?
- Customer relationship: Does the brand receive the minimum lawful order and service data needed to support the buyer and earn repeat business?
Product video now needs a verifiable layer
An agent will usually read structured product information before it “understands” a full commercial the way a person does. Brands should not assume that producing a video automatically makes the claims inside it available to an AI shopping system.
A stronger production package contains two connected layers:
- The human persuasion layer: story, visual identity, use context, emotion, pacing, and motivation.
- The machine-verifiable layer: every material claim mapped to a product field, demonstration shot, transcript segment, condition of use, timestamp, and approved version.
If a video says installation takes three steps, the deliverables should also include three separable demonstration clips, the applicable models, required accessories, exceptions, and a structured instruction. The commercial can persuade a person while the underlying content system gives an agent controlled facts to retrieve.
This changes how production bids should be compared. A cheap master video with no proof shots, transcripts, claim map, or version history may be difficult to reuse across agentic channels. A shoot that delivers a master film plus reusable evidence clips and a structured product-fact pack can cost more upfront and still reduce future reshoots and repeated explanation. More practical thinking on product content systems is available in ONCE’s English insights.
A controlled 30-day pilot
Week 1: Choose one market, one channel, and 10–20 low-risk SKUs. Assign owners for identifiers, variants, price, inventory, shipping, returns, and product claims.
Week 2: Enable Read and Recommend. Test with real customer questions and log variant confusion, missing limitations, and stale facts. Fix each error at its source.
Week 3: Keep checkout on the merchant site. Record recommendation, referral, cart creation, and purchase as separate events. Review contribution margin, cancellations, returns, and service load—not conversion alone.
Week 4: If data freshness, claim accuracy, consent, payment control, and refund ownership pass review, pilot Act on one channel and a small SKU set. Keep high-value purchases behind human confirmation. Do not add Resolve merely to make the journey look complete; post-purchase authority needs its own access model, logs, and escalation design.
Agentic commerce is a controlled delegation system
The upside may come from a shorter path to purchase, or simply from better product discovery. Both are worth testing, but they should not inherit the same risk assumptions.
Our recommendation is sequential: let the agent read before it recommends; prove that recommendations are reliable before it spends the customer’s money; define refunds and customer ownership before automating service. A mature agentic commerce program is not one that opens every door. It is one that knows why each door opens, who holds the key, and who is accountable when the action fails.
Sources and scope
- Shopify Help Center, “Managing agentic storefronts”, accessed October 2, 2026.
- Shopify Help Center, “Shopify Catalog and product discovery for agentic storefronts”, accessed October 2, 2026.
- Google Merchant Center Help, “About the Universal Commerce Protocol (UCP) and UCP-powered checkout feature on Google”, accessed October 2, 2026.
- Stripe, “Agentic Commerce: A Getting Started Guide”, 2026.
- OpenAI Developers, “Products — Agentic Commerce”, accessed October 2, 2026.
- Cloudflare, “Cloudflare Gives AI Agents an Identity and a Wallet”, August 4, 2026.
- The Wall Street Journal, “Retailers From QVC to Kohl’s Are Split on Agentic Shopping”, October 2, 2026.
Platform documentation supports claims about those platforms’ capabilities and plans. It does not establish uniform retailer adoption or guaranteed commercial impact.
Production examples and next steps
Define which images are exploratory and which must preserve a verified product. The AIGC off-road motorcycle concept film is a concept-work reference.
Review AIGC commercial production and Chasing Speed Through the Wild: An AIGC Off-Road Motorcycle Concept Commercial. To discuss your project, send the product references, proposed shots, review criteria and required outputs through the English project enquiry form.
