The short answer
Retailers should create a tiered access policy for AI shopping agents before they choose an integration partner. The policy should distinguish what an agent may read, compare, place in a basket, hand off to a customer and, where a retailer is prepared to support it, purchase with explicit approval. Recent reporting shows that retailers are already taking different routes. That is evidence of a live commercial design question, not proof that one route will outperform the others.
On 2 October, the Wall Street Journal reported that QVC was welcoming agents, Tapestry’s Kate Spade setup allowed agents to browse but not complete purchases, and Amazon blocked them. The same report said Gap, Walmart and Best Buy had publicised Muse partnerships, while Shopify said Muse would be able to buy from merchants on its platform. The point is not to copy a named retailer. It is to make the scope of access a deliberate business decision.
Modi’s view: “AI access” is an unhelpful binary. A retailer needs a commercial permission model: what can an agent see, what can it compare, what can it submit, where must it hand back to a person, and how will the brand know whether the channel created useful demand?
Why a bot block is not a strategy
A conventional bot rule was built for scraping, fraud prevention or unwanted automation. An AI shopping agent may still create those risks, but it may also arrive with a buyer’s stated task: compare a product, explain a return policy, check availability or complete an approved checkout. Treating all automated traffic as the same loses the distinction between an untrusted scraper and a controlled channel with a declared role.
The practical issue is not whether an agent is “good”. It is whether each action has an appropriate contract. Product discovery generally needs different permissions from order modification; a price comparison needs different data from a payment instruction.
| Policy stage | What the agent may do | Commercial purpose | Guardrail to decide first |
|---|---|---|---|
| Discover | Read public product, category and policy facts | Qualify a buyer and explain relevance | Current, crawlable and factual information |
| Compare | Use approved attributes, price and availability | Support a constrained choice | Variant, stock and pricing accuracy |
| Handoff | Send a buyer to a verified checkout or account flow | Preserve consent and identity control | Clear source attribution and session boundaries |
| Approved purchase | Complete a defined transaction under explicit user approval | Reduce friction in a bounded use case | Authentication, payment authority, cancellation and recovery controls |
Build an access ladder, not a vague “agent-ready” claim
A useful starting point is an access ladder. Each step should have an owner, a reason to exist and an observable outcome.
1. Make public commercial facts dependable
An agent cannot reliably answer a product question if the underlying page is incomplete, contradictory or hidden behind an interface it cannot use. Shopify’s agentic-commerce guidance makes the same operational point: agents need complete product attributes, factual descriptions, price, availability, shipping, returns and FAQs. Those are not “AI extras”. They are the commercial facts a human buyer also needs.
Start with the pages that influence a decision: category pages, top products, service descriptions, delivery information, returns and pricing. Use plain language, visible dates where facts change, consistent variants and an accountable publishing owner. This is the foundation of SEO, AEO and GEO: make an approved fact easier to find and verify without inventing a performance claim.
2. Separate observation from comparison
Public discovery can often be the least risky stage, but it should not become an accidental permission to query every database or infer every commercial rule. Decide which attributes are safe to expose and which require a signed-in account, a quote process or a human conversation.
For example, a public product title, compatibility detail and standard return policy may support an agent’s comparison. Contract-specific pricing, customer history, eligibility logic and inventory reservation typically need stronger controls. The policy should say who decides when that boundary changes.
3. Keep checkout as a distinct design decision
Shopify notes that some channels can send a shopper to a store to complete checkout, while others use channel-specific paths. That distinction matters because a discovery experience is not a payment authority model. A retailer can support useful agent-led research without delegating cart changes, payment selection, refunds or post-purchase account actions.
The question for a leadership team is not “can an agent transact?” It is “which transaction, under which verification, with what customer confirmation and with what recovery path?” A clear answer prevents a proof of concept from silently becoming a new payment surface.
4. Define the data exchange before the launch message
If an external channel can access a product feed, order status or customer context, document the minimum data it receives, the purpose, retention expectations, customer notice and revocation path. Do not rely on a platform label to answer these questions. Review the current agreement and technical documentation with the teams accountable for privacy, security, commerce operations and customer service.
For regulated or higher-consideration categories, this should be part of the AI governance work, not an afterthought in a growth experiment.
A 30-day operating test for retailers
The first controlled launch does not need to prove a grand “agentic commerce strategy”. It needs to answer whether a bounded journey is legible, safe and commercially measurable.
| Week | Decision to make | Evidence to inspect | Owner |
|---|---|---|---|
| 1 | Which buyer task is in scope? | Top questions, product complexity, policy gaps | Commercial lead |
| 2 | Which facts are safe for the channel? | Product, pricing, availability and return data | Ecommerce and content owners |
| 3 | Where does the agent stop? | Sign-in, basket, payment and service handoff rules | Product, security and legal owners |
| 4 | What will count as a useful result? | Qualified sessions, assisted conversion, support impact and error cases | Analytics and finance owners |
This is deliberately less exciting than “turn on autonomous checkout”. It is also more likely to reveal where the real work sits: conflicting product information, weak policy pages, unclear ownership, missing attribution or a handoff that loses the buyer.
Measure a channel without inventing incremental revenue
Agentic access can create a new referrer, a new checkout path or an indirect influence on a customer’s shortlisting process. That does not automatically make it incremental revenue. Track the agent or partner source where it is available, compare the quality of handoffs with comparable traffic and retain a visible record of failures: unsupported products, stale price answers, wrong variants and abandoned transfer points.
Use a small set of questions in every review:
- Did the channel send qualified people or merely more visits?
- Did the public facts remain correct through the decision journey?
- Did the handoff preserve consent, identity and attribution?
- Which service issues, returns or support contacts appeared after the agent interaction?
- What did the retailer learn that improves the normal website as well?
A good result is a measured answer to those questions, not a claim that an agent will replace the store, search or customer relationship.
What this means for marketing and merchandising teams
The availability of agents puts more value on the information that survives comparison: clear specifications, current price and inventory context, delivery and returns, compatibility, evidence and a simple way to escalate a question. It makes generic superlatives less useful because a system cannot responsibly compare “best-in-class” language without attributes behind it.
That is why the work belongs across merchandising, performance marketing, SEO, customer service and product. The marketing team may see the traffic first, but it cannot resolve a variant taxonomy, a policy exception or a payment authority decision alone.
For leadership context, Competing in the Age of AI is a useful Amazon UK Associates reading resource for operating-model discussion. It is not evidence for any retailer’s agent capability, commercial result or compliance position. Integrated.Social may earn from qualifying purchases.
The bottom line
Retailers do not have to open every door to an AI shopping agent. They do need to decide which door exists, why it exists and who owns it. A clear access ladder turns an anxious yes-or-no debate into a controlled commercial design: public facts for discovery, approved data for comparison, deliberate handoff for checkout and explicit authority only where the business is ready to support it.
References
- The Wall Street Journal: AI Agents Aim to Change Shopping. Some Retailers Are Locking the Doors, 2 October 2026.
- Shopify: Agentic Commerce: Benefits & How To Get Started, accessed 3 October 2026.
About the Author
Modi Elnadi is the founder of Integrated.Social, a London AI growth consultancy working across B2B, B2C, B2B2C and DTC. Since 2014, he has connected AI search visibility, performance marketing and governed agentic workflows to commercial evidence and qualified demand. His view: automation becomes useful when its permission, source and measurement boundaries are as clear as its promise. Connect with Modi on LinkedIn or explore AI marketing strategy.
