Amazon built its shopping assistant to keep customers from drifting to ChatGPT and Google; on May 27 it began renting that assistant to its competitors. The deal converts a defensive tool into a revenue line: every retailer that runs Amazon’s shopping AI spends its own agentic-commerce budget with Amazon and seats Amazon as infrastructure under its storefront. That is the argument; the rest is mechanism. The Agentic Shopping Assistant on AWS runs on the same engine Amazon credits with nearly $12 billion in incremental sales last year, and Kate Spade is among the first retailers to ship one on it. Amazon is selling something more valuable than a chatbot: the ground every rival’s chatbot now stands on.

The pitch is speed, and speed is the argument. Amazon hands over the architecture, starter code, and its own engineers’ guidance, so a retailer can launch in roughly 60 days rather than the years Amazon says building it alone would take. Kate Spade’s AI Gift Concierge went live on April 13, built on Amazon Bedrock and running Anthropic’s Haiku 4.5. Amazon claims conversational sessions convert at three and a half times the rate of keyword search, which is the figure that makes sitting the transition out look expensive. For a retailer behind on agentic commerce, that speed decides whether a storefront agent ships this quarter or waits for a memo explaining why it hasn’t.

This is the oldest move in Amazon’s book, and its cleanest. Amazon built computing capacity for its own store and rented the surplus as AWS; it built warehouses for its own orders and opened them to sellers as Fulfillment by Amazon; it built an advertising system to sell its own placements and now rents it to the brands on its shelves. The shopping assistant is that maneuver applied to intelligence. Notice where it sits: not in the retail division rivals distrust, but in AWS, the division that already hosts many of their websites. Amazon worked out long ago that owning the rails pays better than owning the store and makes fewer enemies. The marketplace made rivals into plaintiffs; the cloud made them into customers.

The reversal is what makes this more than a new line in the AWS catalog. Amazon built Rufus, renamed Alexa for Shopping in May, to defend the most valuable ground it owns: the moment a shopper decides what to buy. As that decision drifted toward chatbots that give Amazon’s catalog no special place, an assistant that kept it on Amazon.com was insurance against being cut out. Selling that assistant turns the insurance into income, and it hands the buyer Amazon’s accumulated edge: the engine was refined on the questions 300 million shoppers put to Amazon’s own tool and on the answers that closed sales. The rival rents that brain; Amazon keeps a paying tenant and a clearer read on what shoppers everywhere are trying to buy.

The billing is where the strategy stops being a metaphor. Amazon has not named a price, but the assistant runs on Bedrock AgentCore, which charges by consumption: compute per session, memory per stored interaction, tokens per evaluation and guardrail check. A retailer therefore pays Amazon in proportion to how well its own assistant works, because the thing being metered is the thing the retailer wants more of. A license fee gets amortized once and forgotten; this cost doesn’t. It rises with how well a rival’s own assistant converts shoppers into buyers.

Amazon has found a way to be paid for the shoppers it cannot keep.

)

The strongest case against this reading is that a starter kit is not a cage. Amazon ships architecture and code, not a sealed box; the retailer keeps its catalog, its brand voice, and its first-party data. Tapestry’s chief information and digital officer, Yang Lu, calls the deal a shared build: “AWS brought the recipe, but together we built the customization our consumers needed.” If the agent stays portable, and open protocols let a retailer lift it to Google or a neutral cloud, then Amazon has sold a commodity and booked the cloud margin it would have earned anyway — the condition under which this thesis fails. But that asks retailers to later pay for rebuilding what the product exists to spare them. It also confuses the wrapper with the engine: a brand voice can move clouds, while the intelligence that makes it convert improves centrally, on Amazon’s telemetry, across every tenant at once.

Most coverage files this under Amazon joining the agentic-commerce race, which mistakes the game. A race implies Amazon means to own the destination, to beat Google’s cart and OpenAI’s discovery layer for the shopper’s last click; it does not have to. Sold through AWS, the assistant pays Amazon whether the winning storefront is Kate Spade’s, a competitor’s, or a brand Amazon has never heard of. Accenture reckons AI agents could carry close to $3.1 trillion of online spending by 2030, and the money question is who collects a fee on that, whoever wins the sale. Kartik Hosanagar, a Wharton marketing professor, warns that leaning on a platform’s AI can “shift the power balance between tech platforms and retailers.” Amazon is not racing to win the channel; it is arranging to be the toll on it.

The choice this leaves retailers is narrower than the 60-day pitch admits. Renting Amazon’s assistant buys a storefront brain this quarter and defers a cost that is harder to price: the day the conversational agent is the store, and the intelligence running it belongs to the company you compete with. A retailer can rationally decide that speed now beats independence later, and for many the arithmetic will favor speed, because the transition is real and building the equivalent yourself is slow. What a retailer should not do is mistake the decision for a procurement question. It is a positional one. If the agent becomes the storefront, the retailers renting theirs from Amazon will have settled the central question of the next decade of commerce by filing it under software, and Amazon will have charged them for the privilege.