The customer is an agent. Who gets the order?

Earlier this month we wrote that AI agents had started shopping. Last week the two largest retailers in the world gave opposite answers to the obvious next question, which is whether to let them in. Walmart said yes and published the rules. Amazon is in court trying to keep them out. Most brands, distributors and logistics companies will have to make the same call with far less leverage, so it's worth looking at how each side reasoned.

Walmart wrote rules. Amazon filed a lawsuit.

On 21 October Walmart announced an Agent Gateway that lets approved AI agents search its inventory and buy for a customer. Each agent gets its own profile inside the customer's account, a bit like an additional cardholder, so its purchases are kept separate. By default the customer approves every order. They can instead give the agent a spending limit of $50, $100 or $200, and they can switch it off at any time.

The agent pays the same price a person would ("We price products, not people," Walmart says). It never sees card details and can't add a new payment method. If it buys the wrong thing, "Walmart's returns process applies." Agent Profiles go live in about a month, and the gateway opens first as a beta for companies and developers.

Amazon has gone the other way. It blocks outside agents from buying on its site, including Meta's new assistant, and it sued Perplexity over its shopping agent. Amazon won an injunction in March. The appeals court lifted it in August. Amazon filed amended claims in September, and on 13 October Perplexity asked the judge to dismiss them. Meanwhile Amazon's cloud business is selling the technology behind its own shopping assistant to other retailers, saying it drove close to $12 billion in extra sales and can be set up in about 60 days.

The reason for the split is simple. Amazon already wins when agents shop. Six in ten AI-assisted purchases by US consumers recently ended on Amazon. In one test, an agent could buy the same item at the same price with the same delivery from three places. The big marketplace still got 40% of the orders, the department store 22% and the brand's own site 16%. If customers already come to you out of habit, you keep the door shut. If they don't, you open it and set the terms. When Walmart's EVP of AI, Daniel Danker, was asked whether this means handing customers to AI platforms, his answer was one word: "intermediation."

Where investors think the money is

The investors we read are not backing chat interfaces. They're backing what sits underneath.

Alex Rampell at Andreessen Horowitz has said the near-term opportunity is infrastructure: payments that work across sites and apps, product catalogues that agents can read, and new ways to track what drove a sale. His colleague Justine Moore says last-click attribution "largely doesn't work in this new world". Rampell also thinks Amazon "could end up growing even faster", which doesn't contradict any of this.

On payments, a16z's Noah Levine argued in March that cards will carry most agent purchases from existing merchants. Visa has issued more than 16 billion tokens, and Mastercard's agent payments are live for US cardholders. New payment rails will mostly serve sellers that card processors can't underwrite yet, like software that charges a fraction of a cent per request. The checkout protocol Stripe and OpenAI built together already has Etsy live, with over a million Shopify stores set to follow.

Greg Isenberg, who invests in AI-native companies, wrote this month that the strongest teams treat the AI model as replaceable. One founder told him "the model is maybe 20% of the product". In that founder's view the rest is industry data, evaluations written by experts, permissions and a clean screen for a person to review. Isenberg also noted that every company wants to connect its data to agents, and "then they look at the data and realise it's a mess." Fixing that is where he sees the work.

That matches what we see with clients. Getting an agent running is the quick part. Most of the time goes into cleaning the data and agreeing who is allowed to approve what.

Shopify made the decision for its merchants

Since December Shopify has treated every store as ready for agents. One setting puts a store's catalogue into ChatGPT, Perplexity and Copilot. The merchant stays the seller of record and sees each order in the admin, tagged with the AI channel it came from. In September Shopify added Meta's assistant as a channel, turned on direct checkout for eligible stores in the US, Canada and Mexico, and made Shop Pay the wallet inside the assistant. Its AI-driven traffic and orders each tripled year on year in the second quarter. Shopify also says AI search that reads a structured catalogue converts at twice the rate of AI search working from scraped web pages.

Shopify's president, Harley Finkelstein, put it this way: "Structured data, trusted payments and deterministic APIs are what actually work." In other words, a clean product feed matters more than the campaign.

Logistics got there first

Retailers are still arguing about whether to let agents in. Logistics companies have been letting them make decisions for a while. DHL's Logistics Trend Radar, published on 24 September, lists agentic AI and AI commerce among the forces that will shape the next five years. It describes agents that spot a likely stock-out, reroute a shipment or adjust capacity before anyone has read the alert. "The next chapter is about action," says DHL's head of trend research, with people kept in charge of oversight.

That matters most for distributors. Once a retailer's reordering agent can compare three suppliers on stock, lead time and price in a second, a distributor that only moves boxes gets squeezed. The order goes to the one whose system gives the agent a correct answer at 10pm on a Saturday.

What to do if you're not Walmart

Three things, starting with the most urgent.

First, treat your product data as your shop window. That means titles, options, stock levels, delivery dates and a returns policy written in plain language. An agent can't make sense of a PDF spec sheet or a "usually in stock" label. If Shopify's numbers hold, leaving this messy costs you about half your agent conversions.

Second, decide your rules for agents before someone else's agent decides them for you. Walmart's version is a good starting point for any business. Each agent gets a named profile and a spending limit, sees the same price as a person, never handles card details, and is covered by your returns policy even when it gets things wrong. Write those five points down this month and add them to your terms.

Third, use an agent on your own front line before you worry about anyone else's. Good places to start are orders that come in on WhatsApp, weekend enquiries and repeat orders from existing accounts. Keep a person approving what goes out, and write down the number it has to move before you start. That's where most of our clients will meet agentic commerce, not in a US chat app. South African online retail is on track for about R159 billion this year, and most of it is bought on a phone.

Where to start

Ask four AI assistants to recommend and buy your best-selling product, and note every mistake they make. That list is your first month of work. It's worth doing whether or not agents arrive on schedule, because every fix also makes it easier for a person to buy.

If you'd like help working through the list, book a free 15-minute audit. That's where every Automaiq engagement starts. Bring the product and the four transcripts. Proof Sprints and Core Builds come with the same guarantee for every client: Measurable ROI in 90 days — or 50% of your project fee back.