When Your Customer Sends an Agent

Picture a customer who wants a coffee machine. It has to fit under the kitchen cabinets, arrive by Friday and cost less than $300, and they want to see the recommendation before anything is ordered. So they hand the job to an AI agent. One retailer lists the machine at $320. Another has it for $295. The agent compares public prices, finds otherwise equivalent terms and recommends the $295 option.

It did a reasonable comparison and still gave bad advice. This customer belongs to the first retailer's loyalty program and qualifies for a $280 member price. The retailer knows that, and the customer earned it. The agent never saw it, because the public page didn't show it and nothing connected the customer's membership to the software acting for them.

That is the problem I want retailers thinking about. Making agent checkout reliable across more retailers and situations is a problem the platforms are pouring money into, and I expect they'll solve it. The harder question is whether the relationship a retailer has built with a customer still counts when that customer sends software to shop for them. For most retailers today, I suspect the honest answer is no.

In September, Meta's Muse agent started checking out with Stripe's Link once the person approves the purchase [1]. Google says its Universal Commerce Protocol already supports direct checkout for hundreds of thousands of brands and retailers across Search, including AI Mode, and the Gemini app [2]. Amazon's Alexa for Shopping will buy an item with your default payment method when it drops to a target price you set [3]. Since March, millions of Shopify merchants have been discoverable in ChatGPT, with the purchase completed through the merchant's own checkout [4]. The details differ by platform, but in each case more of the comparing, and some of the buying, happens before a person reaches the retailer's site.

Bringing the membership along

The fix starts with letting customers bring their membership with them, on their terms. UCP's Identity Linking does this with ordinary OAuth. The customer authorizes the agent platform to act for them with a specific retailer, and the retailer decides which operations need that linked identity, with permissions scoped to the work involved [5]. Someone who would rather shop anonymously still can. I expect most agents to arrive without a linked account for a while, which is why linking has to be worth the customer's trouble. As the spec puts it, identity linking upgrades the experience; it does not gate it.

Once the account is linked, UCP's Loyalty extension carries what the agent was missing in my example: memberships and tiers, member pricing and free shipping applied in real time, and a preview of what the purchase would earn [6]. The retailer returns the $280 price that applies to this customer without handing over their purchase history.

Say the purchase would also earn 560 points. The agent may well know roughly what those points are worth, because the program publishes a value or the retailer supplies one. They still aren't money off this purchase. They're still part of the decision, and one more reason for the agent to pick the retailer the customer already has a relationship with. UCP also keeps earning and spending apart. The loyalty extension covers the points you would earn. Spending a balance you already have is treated as a form of payment, or left for a future redemption feature [6]. That's the right call, because spending a balance carries risks that earning one doesn't.

From the customer's side, none of this is complicated. They expect "You know I'm a member. Does that help here?" to have an answer, and it should have one when their agent asks too.

Finding the offer isn't permission to buy it

Our customer also said they wanted to see the recommendation before anything was ordered. That's easy to honor in a conversation and much harder to preserve once the work moves between an agent platform, a retailer and a payment provider. Permission to see member pricing is not permission to use a saved card.

AP2, the Agent Payments Protocol, carries verifiable evidence of what the customer actually authorized. In UCP's version, the retailer signs the checkout terms, and the platform wraps that signed checkout in a mandate once the customer approves. Before completing the order, the retailer confirms that the terms inside the mandate still match the session, totals and line items included [7]. If the price moved, the earlier approval doesn't cover the new transaction. An agent's confident summary can't make a stale price current, and the protocol doesn't ask anyone to trust that it would.

AP2 also separates two kinds of approval [8]. A closed mandate is a signed record of one specific checkout. An open mandate is a set of constraints the customer signs in advance, such as a budget and conditions, which lets the agent approve a matching checkout without coming back to them. The spec recommends that open mandates expire as soon as the task allows. I like that the default is narrow, because people want different amounts of automation. Some want the agent to finish the job. Others want help choosing and then a moment to look at the details themselves. When a purchase falls outside what was approved, a good handoff brings the current offer and the work already done back to the person instead of starting over.

Marketing consent is a separate decision again. UCP's Buyer Consent extension lets a platform pass along the customer's choices about marketing email and SMS, analytics and data sharing [9]. Agreeing to hear from a retailer is not agreeing to buy the coffee machine, and the systems involved shouldn't treat them as one decision.

Who shapes the choice

Here is the uncomfortable part for retailers. If an agent platform builds the shortlist and decides which differences are worth mentioning, it owns much of the shopping conversation. Keeping checkout on your own site, as Shopify merchants do in ChatGPT, doesn't mean you kept the discovery relationship.

Platforms that own discovery tend to charge for it. Retailers with long memories will recall October 2012, when Google Shopping in the U.S. stopped showing free listings and began showing only merchants paying for Product Listing Ads [10]. The new platforms are already building their commercial layer. OpenAI started testing ads in ChatGPT in February [11], and it is piloting ad campaigns that retailers build directly from their product feeds [12]. OpenAI says ads don't influence ChatGPT's answers and are always separate and clearly labeled [13]. I take that at face value. A sponsored product sitting beside a recommendation still competes with it, though, and the retailer paying for placement has an obvious advantage over the one that only has a better price for its members.

So the argument that platforms will surface member pricing because good advice is in their interest only goes so far. An agent that recommends $295 when its user could pay $280 has given worse advice, and an assistant competing on the quality of its help has a reason to care about that. It also sells ad space next to those answers, and nothing in UCP or ACP decides what sits beside your member price.

That still doesn't make withholding benefits a sensible strategy. A retailer that keeps member pricing away from agents has already lost the comparison it was best placed to win. One that exposes it may lose some of the customer's attention to paid competitors, but it gives the agent the facts it needs to recommend the retailer, and it gives the customer a reason to link their account in the first place. Making the relationship harder for an agent to reach is an odd way to protect it.

Knowing who's shopping

I've argued before that enterprise software should work with whatever AI the customer chooses [14], and retail is where that gets tested with real money. At Exchange Solutions, working out how ES Engage™ should interoperate with agents is now a product priority, and the starting point is clear: you can't tailor an offer to a session when you can't tell whether a person or an agent is in it. ES Engage decides, during the session, whether a shopper should see an offer, and it reads intent from clicks, time on page and cart events. An agent driving the page produces clicks, but not the dwell time, hesitation and browsing your models read as interest. An agent calling tools changes the cart with no clicks at all. And since August 5, Shopify has had WebMCP tools switched on for every Liquid storefront, with nothing for the merchant to install. It's limited to Chromium browsers during an origin trial for now, but agents can already search, fill a cart and head to checkout on those stores without the merchant making a single decision [15].

So retailers should make detection their first priority. Most can't say today what share of their sessions are agents, and that number is how you size the opportunity. Flag agent and tool-driven sessions, report them as their own segment while keeping them in the test and control groups you use to judge whether an offer changed anyone's behavior, and count the orders that arrive with no human session behind them. Left unflagged, they disappear into the average. An overall lift figure won't tell you whether an offer works differently when an agent is shopping, and an agent that starts each visit from a fresh browser can land in both groups.

Once you can see agents, work out how to engage them, and think wider than offers. An agent can carry the loyalty benefits a customer has already earned, and it can bring you new members. If our coffee machine shopper weren't a member yet, the most useful thing the agent could say is that joining gets them the $280 price, and then, with their approval, join on their behalf. Today's protocols don't make that easy.

UCP's loyalty extension is built for existing members and doesn't define how someone joins [6], so a retailer that wants agents to enroll customers has to expose that itself, as a tool the agent can call once the customer agrees to the program's terms. Whatever you expose, measure whether it earns incremental margin or just gives money away. Some agents will be price hunters that collect every discount available for a customer who was going to buy anyway, and an offer that only attracts them is a subsidy, not marketing.

We're treating this as a priority across ES Engage and ES Loyalty™'s ecommerce experiences, but it isn't a vendor problem. Every retailer selling online should start by learning to see agents in its traffic, and shouldn't wait for perfect detection before giving them a way in. When you get to engagement, hold on to one distinction. A member price the customer earned, delivered through an account they chose to link, is the relationship. A promotion handed to any anonymous agent that asks can be leakage. Get the first in front of agents, and be much more calculated about the second.

Measuring whether the relationship gets through is the last piece, and it's harder than it sounds. The retailer never sees what the agent finally tells its user, so the measure has to come from what the retailer does see. How many agent sessions arrive with a linked identity? When a linked customer qualifies for a benefit and we return it, how often does that session end in a checkout, compared with an equivalent journey on the website? A persistent gap won't prove the relationship got dropped, but it tells you exactly where to start digging.

The website still matters

None of this replaces the website. People will still want to look closely, ask questions and take over, and early browser work such as WebMCP, still a draft rather than a standard, is aimed at that shared moment where a person and their agent work on the same page [16]. I just don't think visiting the site should be the price of being recognized as a customer. If someone chooses to send an agent ahead of them, the relationship they've built with you should count there too.

References

1. Meta, Introducing Muse, September 8, 2026
2. Google, Boost your holiday sales with these agentic commerce updates, September 16, 2026
3. Amazon, AI features to help you find the best deals, accessed October 23, 2026
4. Shopify, Millions of merchants can sell in AI chats, March 24, 2026
5. UCP, Identity Linking
6. UCP, Loyalty Extension
7. UCP, AP2 Mandates Extension
8. AP2, Specification: Authorization Modes
9. UCP, Buyer Consent Extension
10. TechCrunch, Google Shopping's Transition To Pay-For-Play Completes On Oct. 17, October 3, 2012
11. OpenAI, Testing ads in ChatGPT, February 9, 2026
12. OpenAI Help Center, Create Campaigns from Product Feeds, accessed October 23, 2026
13. OpenAI, Our approach to advertising and expanding access to ChatGPT, January 16, 2026
14. Shane O'Neil, Your AI Should Be Able to Use Our Software, Exchange Solutions
15. Shopify, WebMCP support for Liquid and Hydrogen storefronts, August 5, 2026
16. Web Machine Learning Community Group, WebMCP Draft Community Group Report, October 20, 2026