The agent found it. Who gets to sell it?
AI-generated image
Hi,
the launch date is nearly locked. When it is, you'll hear it from me here. It's a matter of days, not weeks.
OPENER
On the evening of 20 September, Amazon blocked Muse, Meta's new personal AI agent, from shopping in its store. Muse books, calls and messages for you, and it also buys. The next day, Shopify announced the opposite: Muse can shop in every Shopify store and pay with Shop Pay.
That reads like closed against open. Look closer, and all three companies open the door exactly where it pays them.
The question underneath: who gets to buy on your customer's behalf? Below is the agent purchase, step by step. Every step a retailer hands over is a step where someone else watches the customer decide, and decides what they see.
THE AGENT PURCHASE, TAKEN APART
1. Discovery: the agent finds the product
This part already works. Adobe expects traffic from AI assistants to US retail sites to grow 130% this holiday season, and 60% of US shoppers plan to use AI tools, up from 47% last year (ICSC). Muse topped Apple's US App Store a week after launch; TikTok added a Buy Direct function and a Shopping Assistant on 5 October.
Source: McKinsey
2. Access: who lets the agent in
Amazon has a point. Muse doesn't identify itself as an agent and appears to store customers' logins. A retailer needs to know who is in its systems and who pays for a mistake.
But the rule runs one way. Amazon's own agent, Buy for Me, shops on other brands' websites unless they object. Outside agents in Amazon's store need permission, which Amazon refuses. It has sat on the board of the Universal Commerce Protocol, the new open standard for agent checkout, since April, yet its marketplace stays closed. The money explains a lot: $68.6 billion in ad revenue in 2025, mostly selling visibility at the moment of purchase. An agent ignores sponsored listings.
The other side isn't charity either. Shopify's door opens on Shopify's terms: one default setting has shared merchants' products with Meta since Muse launched, and the agent pays through Shop Pay. A different toll, but a toll. Meta itself shut general AI chatbots out of WhatsApp Business last year. Here it seems like two advertising companies are fighting over who controls the moment of decision.
3. Checkout: who sells, and who learns
The least-used step so far. OpenAI pulled Instant Checkout from ChatGPT this year with around 30 Shopify merchants live, and Forrester finds buying inside an AI assistant is the rarest thing people do there. The Universal Commerce Protocol, built by Shopify and Google, keeps the retailer as seller of record. TikTok's new one-click Buy Direct runs on it, and Zalando and Adyen back it.
Seller of record isn't owning the journey, though. On its own site, a retailer decides navigation, related products, offers, and the path to checkout. Inside a feed or an agent, the platform makes part of those calls. Buy on TikTok after asking its assistant, and TikTok sees which video sparked the interest, which questions came up, and what tipped the decision. The retailer gets the order. The platform gets the reasons.
4. Identity and payment: who vouches, who pays
Visa and Mastercard are building "Know Your Agent": signatures that show which agent is acting, for which customer, and whether it may only research or also buy. On 30 September, Mastercard began testing a score for whether an agent started a transaction. That settles who is buying, not the rest:
Disputes: if a customer says "my agent bought that, not me", nobody knows what the merchant must prove. 93% of merchants say the agent provider should pay (PYMNTS).
Fraud signals: device, location, and typing behaviour vanish when an agent buys.
Fees: Meta plans to take a small cut of Muse transactions. Every agent in between is a new toll.
5. Recognition: does the agent know your customer?
An agent that can find and pay you still needs to know who it's buying for: member price, points, saved size. That needs the customer's account linked to the agent, and Google lists loyalty among its protocol's next features. So, a loyalty programme the agent can't see doesn't exist for the agent.
WHERE I LAND
Amazon can say no because it needs no outside agent for any of these steps: product searches start in its search bar, and it has its own assistants, checkout, and customer cards. For almost everyone else, blocking agents means being invisible at step one.
Every agent purchase needs two yeses. The customer has to send the agent, and the store has to let it in. A retailer only controls its own yes, and only partly: it decides which agents get its product data, and whether they can check out themselves or must hand the customer back. On Shopify that's a switch per channel; elsewhere it's product feeds and bot settings. Agents that browse like a person get past both, which is exactly what Amazon objects to. Know Your Agent would make them visible, but visible also means easy to block, so few agents will volunteer an ID yet.
Within that, I'd write product data for the questions people actually ask. Take an invented example: a shopper tells their agent "shirt for an outdoor wedding in Barcelona in July, 32 degrees, can't look creased after the car ride". "100% linen, regular fit" won't win that answer; a line about how it breathes above 30 degrees and holds its shape after an hour sitting down might. I'd make loyalty prices readable for agents, keep checkout and the customer record where I can, and hold back on agent payments until liability is settled.
What I don't know is whether agent checkout converts. Otto says baskets built with its own assistant are 40% bigger than from normal search. Walmart reportedly sees in-chat purchases convert at a third of its site rate. If Walmart is typical, the real fight stays at step one.
Sources: Bloomberg, GeekWire, Forrester, Mastercard, Google
ONE QUESTION
Which one did you pick, and why? Hit reply. I read every one.
Talk soon, Luca
