Picture the first Monday after Thanksgiving. Treat it as an illustration, not a report. Your ecommerce lead notices a cluster of orders that arrived within seconds of one another, from a handful of user agents, with item mixes no person would assemble on purpose. Some look like real customers who asked an AI assistant to shop for them. A few look like a bot that misread its instructions. By Wednesday the first return request arrives, and the complaint is simple: the assistant ordered the wrong size, and nobody approved it. Who pays for that?
Most mid-market retailers and consumer brands have not decided. This fall, the two largest retailers in the country decided for themselves, and they chose opposite answers. Amazon has been closing its doors to outside shopping agents. Walmart has built a door for them to walk through. The gap between those choices is now yours to fill, and the holiday rush is a poor time to fill it by improvising.
Two Giants, Two Answers
In late September, Amazon began blocking Meta’s new AI agent, Muse, from its retail site after Meta declined to pull it, according to Bloomberg reporting carried by The Star. An Amazon spokesperson said the company had “requested that Meta remove Amazon from the experience.” The move fit a pattern. Earlier this year, as FashionUnited reported, a federal court issued a temporary order barring Perplexity’s Comet browser agent from buying on Amazon, and the same report noted that Amazon now requires AI agents to identify themselves. When TechCrunch surveyed the landscape on October 21, its reading was plain. Amazon has started blocking agents that try to buy on a shopper’s behalf.
That same day, Walmart went the other way. CEO John Furner announced Agent Gateway at the Masters of Scale Summit, as reported by TechCrunch and AI Insider. Agent Gateway is an MCP server, a standard front door that lets approved agents search inventory and complete purchases for a customer at Walmart and Sam’s Club. We have written about why that kind of entrance matters in our piece on MCP agent docks. The short version is that one defined door is far easier to govern than a thousand improvised ones.
The welcome comes with house rules. Per both reports, Walmart plans to introduce Agent Profiles in November. Each agent would get its own profile inside the customer’s account, so its purchases appear apart from the customer’s own. The customer could require approval for every purchase or set a spending limit, an arrangement TechCrunch says Walmart likens to adding an authorized user to a credit card. Agents are not supposed to act beyond what the shopper intended, see payment details, or touch more data than the task requires. Agents that break the terms get blocked. And Walmart’s normal returns process applies even when an agent does not do what the shopper meant.
That last point deserves a second reading. Walmart has said, in public, that it will absorb the cost of an agent’s mistake through its standard returns process. It is a deliberate business choice. You have probably not made yours yet.
A word on what is settled and what is not. Agent Profiles were described as arriving next month; as of this writing they are not live, and the details may shift when they do. Our account of Walmart’s approach rests on reporting of Furner’s remarks rather than on terms Walmart has published itself. Treat it as a strong signal of direction, not a finished rulebook.
Why This Lands on Your Desk Now
The demand side is not waiting for the policy side. In a survey released in August, the returns technology vendor Narvar reported that 65% of U.S. consumers plan to use AI for at least one part of their holiday shopping, while only 8% of the 100 retail decision makers it surveyed said they were very confident using AI to improve the shopping experience. Those are a vendor’s numbers, and the retailer sample is small. The direction is still hard to miss. Shoppers are moving faster than the stores they shop.
Most of that AI use today is research rather than buying. A shopper asks for gift ideas, compares prices, reads summaries of reviews. But the buying arrives through the same tools, and agents already reach your site whether you planned for them or not. If you have no policy, you still have one. It is simply written by accident. Your bot filter, your fraud tool, or your hosting vendor makes the call on each visit, and none of them knows your margins, your returns budget, or which customers you would hate to turn away.
If you have no policy, you still have one. It is simply written by accident.
Consider what that looks like in practice, again as an illustration. A specialty outdoor retailer with a strong loyalty program runs a standard bot management service. The service sees fast, scripted traffic and blocks it, which is exactly what it was built to do. Some of that traffic is a scraper copying prices. Some of it is a longtime customer’s assistant trying to reorder the same boots in a new color. The retailer never chose to turn that customer away. A vendor default chose for it, quietly, and nobody in the building knows it happened.
The Three Doors
Every retailer and brand will land somewhere among three choices. None of them is wrong. Each one has a price.
The first door is to block, which is the Amazon path. It protects your customer relationship, your on-site advertising and search revenue, and your data. It makes the most sense when your own app or loyalty program is the moat, and when customers come to you directly out of habit. The cost is that a shopper who lives inside an AI assistant may simply buy from someone who lets the assistant in. Amazon can afford that bet because its pull on shoppers is enormous. Most mid-market sellers have far less, and should be honest with themselves about it.
The second door is to allow on terms, which is the Walmart path. You let named, approved agents in through a defined route, and you set the rules: who they are, what they can see, what they can spend, and who owns the mistake. This is the door most likely to capture new demand. It also takes the most work, because terms you cannot enforce are only a hope. You need a way to recognize the agent, a way to cap it, and a way to cut it off.
The third door is to rate limit and watch. You let agents in slowly, cap how fast and how much they can buy, and study what they actually do before you commit to more. For many mid-market firms this is the honest answer for this holiday season. It buys time without slamming the door. It also produces something you do not have today, which is evidence.
Brands that sell through others face a split version of the question. If your goods are listed on Walmart.com, part of this choice has already been made for you on that channel, under Walmart’s terms. Your own site and your direct-to-consumer program still need an answer, and it does not have to match. Shopping-center owners will meet the question through their tenants, and through their own mall apps, gift card programs, and loyalty data. How agents decide which sellers to trust in the first place is its own subject, one we explored in our piece on buyer agents scoring vendors.
Who Owns the Return
This is the part most policies skip, and it is where the money sits.
This is the part most policies skip, and it is where the money sits.
When an agent orders the wrong size, the wrong color, or ten of something instead of one, three parties could carry the cost. There is the shopper who sent the agent, the company that built it, and you. Today, with no rules in place, it will almost always be you. The customer will call your service line, your representative will follow the standard script, and your normal return policy will do what it always does.
You can choose to accept that, as Walmart has. It is a fair choice if agent orders turn out to be large, frequent, and loyal, because then the returns are simply a cost of acquiring a valuable new kind of customer. But make the choice on purpose, and put a number on it. Or you can set different terms for agent orders. You might require a human approval above a set dollar amount. You might exclude final-sale and custom items from agent purchase entirely. You might offer a shorter return window for agent orders, or require a clear record that the customer approved the purchase before it shipped.
Whichever way you go, the decision has to show up in more than one place. Your terms of service, your customer service script, your payment setup, and your fraud rules all need to say the same thing. A policy that lives in the terms of service but not in the call center script will be overridden on the first angry phone call. That is not a failure of the person on the line. It is a failure of alignment above them.
The One-Page Policy
Your next step is a one-page agent access policy, signed by the CEO, before peak season begins. It should answer five questions, and each deserves a sentence or two, not a chapter.
Start with the door: block, allow on terms, or rate limit, and whether the answer differs by channel. Then settle identity, meaning how you will tell a real agent acting for a real customer from a scraper or a bot, and what you will do with agents that refuse to identify themselves. Then set spending, which means a default cap per agent order and the point at which a human must approve. Then decide returns: who pays when an agent gets it wrong, and whether agent orders follow your standard policy or a separate one. Finish with the off switch, which names who can block an agent in the middle of the season and how quickly that can happen.
The off switch matters more than it looks. In a busy December, a misbehaving agent can place a great many orders before anyone in a weekly meeting notices. If the authority to stop it sits with someone who is on vacation, or requires a change ticket that waits two days, you do not really have an off switch. Name a person. Name a backup. Test it once before Thanksgiving.
Keep the whole thing to one page. A longer document will feel more thorough and will be read less. Revisit it in January with real traffic and order data, when you will know far more than you do now.
What This Will Not Do
A policy will not stop every bad bot, and it will not tell you which AI assistants your customers will favor in 2027. The ground is moving. The Perplexity dispute with Amazon is not settled, Walmart’s profiles are not live yet, and some of what you write this month will need to change by spring. That is acceptable. A written choice you revise is better than an unwritten one your vendors make for you, one request at a time.
It also will not replace judgment about your own customers. A luxury brand whose clients value a personal touch may reasonably decide that agents are not welcome yet. A replenishment business selling the same household goods every month may decide the opposite. The policy’s job is to make that judgment visible and deliberate. It does not make the judgment for you.
Three Questions for Your Next Leadership Meeting
Ask your ecommerce and technology leads three things. How much of our traffic and order volume comes from AI agents today, and how do we know? If an agent buys the wrong thing this December, what happens, step by step, and who pays? And who in this company has the authority to let an agent in, or shut one out?
Listen closely to the second answer. If it begins with a guess, or it changes depending on who is speaking, you have found the gap.
If the answers are vague, you are not alone, and you are not late yet. An Innovation Vista technology and AI assessment starts with exactly these questions, drawing on a bench of more than 450 former CIOs, CTOs, and CISOs, many of whom have carried retail systems through a peak season. The answer you reach may be to block, to welcome, or to wait and watch. What matters is that you chose it.
The answer you reach may be to block, to welcome, or to wait and watch. What matters is that you chose it.
