Amazon Will Sell You Its Shopping Agent: Should You Buy?

The AWS Agentic Shopping Assistant takes the conversational AI that powers Amazon's own storefront and packages it — architecture, starter code, and AWS professional services — for any retailer to deploy. AWS announced it on May 27, 2026, framing the build as "weeks rather than years." For eCommerce operators, that promise is real, and so is the catch underneath it.

The headline coverage fixated on two vendor figures: a 3.5× conversion lift for conversational shopping and a roughly 60-day deployment window. Both are Amazon's numbers, and both deserve scrutiny. The more useful question for a merchant is the one almost nobody asked — what has to be true about your product data before an agent can deliver any of that lift on your site.

This guide treats merchant readiness as the actual decision. We break down what shipped, the Kate Spade reference build on Anthropic's Haiku 4.5, the difference between the packaged ASA and the AgentCore platform underneath it, the unit economics that make consumer-scale agents affordable, and the strategic cost of running your conversational commerce on the infrastructure of the company you compete with. Two proprietary tools — a build-vs-buy matrix by retailer archetype and a catalog readiness scoring card — turn the abstract debate into something you can run against your own feed.

On May 27, 2026, AWS announced the Agentic Shopping Assistant (ASA), a packaged offering that gives retailers the architecture, starter code, and expert guidance drawn from Amazon's own shopping assistant. Two weeks earlier, on May 13, Amazon had renamed that assistant from Rufus to Alexa for Shopping — the same underlying technology AWS is now licensing to outsiders. According to Amazon, Rufus reached more than 300 million customers during its beta.

AWS's pitch is speed. Rather than assembling a conversational commerce stack from raw components, a retailer adopts a blueprint that AWS says compresses the work to roughly 60 days — "weeks rather than years." That timeline is vendor-stated, and it assumes the hardest precondition is already met: a product catalog structured enough for a language model to reason over. More on that gate in section 04.

Treat the launch claims carefully. Amazon states its shopping assistant drove on the order of $12 billion in incremental sales and that conversational sessions convert at 3.5× the rate of traditional keyword search — but both are Amazon's own figures, observed on Amazon.com, with no independent audit found. They are a reason to pay attention, not a forecast for your storefront. A separate, independent McKinsey figure puts AI-generated product recommendations at 4.4× higher conversion than traditional search; it is a different measurement and should not be conflated with Amazon's 3.5× claim.

The launch's anchor case is the Kate Spade AI Gift Concierge, built by Tapestry — Kate Spade's parent company — on Amazon Bedrock AgentCore. It went live on April 13, 2026, making it the first production deployment of this stack outside Amazon itself. (Kate Spade New York is the brand; Tapestry, Inc. is the parent, and Digital Commerce 360 ranks Tapestry #40 in its Top 2000 — with Amazon at #1.)

The agent was purpose-built around a specific pain point: Amazon cites research that 53% of shoppers report stress during gift purchases. The Gift Concierge turns that into a guided conversation — occasion, recipient, budget, style — rather than a keyword search. Under the hood it runs on Anthropic's Claude Haiku 4.5 (API ID claude-haiku-4-5-20251001) via Amazon Bedrock, and Tapestry completed roughly 2.5 months of testing before the customer-facing launch.

That phrasing — "AWS brought the recipe" — is the honest shape of the offering. ASA is a blueprint plus services, not a turnkey product you flip on. The customization Tapestry references (brand voice, the gift-finding flow, the curation logic) is where the real work lives, and it is work only a retailer with a clean, queryable catalog can do well. The roughly 2.5-month Kate Spade testing window is also a useful reality check against the headline 60-day figure: a recognized brand with deep resources still spent the better part of a quarter getting it right.

The single most common confusion in launch coverage is treating the ASA and Amazon Bedrock AgentCore as the same thing. They are layered. AgentCore is the platform — a modular agent runtime. ASA is the packaged offeringbuilt on top of it, with Amazon's shopping-specific scaffolding and AWS professional services wrapped around. You can adopt AgentCore without buying ASA; you cannot buy ASA without running on AgentCore.

AgentCore ships as a set of modular components — Runtime, Memory, Gateway, Identity, Code Interpreter, Browser, Observability, Payments, Evaluations, and Policy among them — and it is deliberately framework-agnostic. It supports LangChain, LangGraph, LlamaIndex, CrewAI, the OpenAI Agents SDK, the Claude Agent SDK, Google ADK, and Strands Agents, and it works with models from Anthropic, OpenAI, Google, Amazon (Nova), Meta (Llama), and Mistral. That breadth matters for the build-vs-buy decision: the same runtime can host a fully self-built agent, not just Amazon's packaged one.

Here is the part the launch coverage skipped. A conversational agent reasons over your product data, so its ceiling is set by how structured and complete that data is. If your feed is full of thin titles, missing specs, stale availability, and no fit or compatibility fields, no amount of model quality rescues it — the agent simply has nothing reliable to reason about.

The readiness gap is wide. Per a Mirakl partner survey, 40% of eCommerce businesses were still standardizing product pages for agentic AI in early 2026, and 33% had not started at all — roughly three in four retailers not yet ready to capture the lift an agent promises. The investment decision a merchant actually faces is not "buy the agent or not." It is "fund the catalog remediation that lets any agent work." That work — building a machine-readable product catalog — is the real line item.

Independent data suggests the upside is real for those who close the gap. Adobe reported that AI-referred visitors to US retail sites converted 42% better in March 2026 than in March 2025, reversing a prior deficit, and that those visitors spent 48% longer on site. A widely circulated industry figure holds that catalogs near full required-field completeness see several times higher visibility in AI recommendation surfaces than sparse ones; we treat that as directional, because its primary research source is not a recognized independent firm. The direction of travel is consistent even where the exact multiple is not verifiable: complete, structured data wins in agent-mediated discovery.

Amazon's model choice for the Kate Spade reference build is a quiet but important signal. It used Anthropic's least expensive flagship — Haiku 4.5 — not Sonnet or Opus. That is a deliberate decision about unit economics. At consumer scale, a shopping conversation might run five to ten turns at a couple thousand tokens each. On a model priced like Haiku 4.5, the model-inference cost of an entire conversation lands in fractions of a cent. That is the reason a 300-million-customer assistant can exist without the inference bill swallowing the margin.

The infrastructure underneath it follows the same logic. AgentCore uses consumption-based pricing with no upfront commitment. Per AWS's published rates, Runtime is billed at $0.0895 per vCPU-hour plus $0.00945 per GB-hour — and critically, only while the agent is active, so the idle time spent waiting on a model response is not billed. Memory runs at $0.25 per 1,000 short-term events and $0.75 per 1,000 long-term records per month. None of those line items is what determines whether the project pays off.

The chart is the whole argument. The running costs of a conversational commerce agent on this stack are genuinely small. The expensive part — for the roughly three in four retailers not yet ready — is the catalog work that has to happen first. Any honest cost model for "should I buy this agent" puts data remediation, not inference, at the top of the ledger. One independent estimate pegged a moderate-traffic support agent (10,000 conversations a month, five turns each) at roughly $50–$200/month in AgentCore infrastructure plus $200–$800 in model inference; that is a third-party figure, not AWS-stated, but it underscores how modest the operating side is relative to the data project.

Strip away the technology and a strategic tension is left in plain view. Amazon is the #1 US retailer. Every retailer that adopts ASA on AWS pays Amazon for the infrastructure behind their own conversational storefront — and runs that storefront on a platform operated by the company they compete with for the same customers. The Next Web put the paradox crisply: the same company that competes with Kate Spade for handbag sales is now providing the AI engine behind Kate Spade's shopping experience.

The ledger is concrete. On one side: the time savings AWS advertises — weeks rather than years — and a managed, battle-tested stack. On the other: recurring AWS infrastructure revenue flowing to a competitor, and the harder-to-price question of what operational and behavioral signals a retailer's agents generate on Amazon's platform over time. None of that makes ASA the wrong choice. It makes it a choice that should be made with the trade-off named, not ignored.

Hosanagar's warning is the forward-looking version of the paradox. As more discovery and purchase shift into agent-mediated channels, the layer that controls the agent accrues the leverage in the relationship — and the retailer that outsources that layer wholesale risks ceding the customer relationship itself. The defensive move is not to refuse agents; it is to keep ownership of the asset agents depend on — your structured catalog and your brand voice — even when the runtime is rented. That is the difference between using Amazon's infrastructure and becoming a line in Amazon's.

Most coverage frames this as a binary: buy the Amazon agent or don't. The honest version has three tiers and depends on who you are. Tier A is buying ASA on AWS — fastest, most managed, most Amazon-dependent. Tier B is self-building on AgentCore with your own Bedrock models — more control, still on AWS. Tier C is self-building on open frameworks (LangGraph and the like) — maximum control and the least platform lock-in, at the cost of the most engineering. The matrix below maps the sensible default for each archetype. For a deeper comparison of how AWS compares to building on open-source agent frameworks, see our Q2 2026 platform breakdown.

The pattern: the more directly you compete with Amazon and the more sensitive your demand signals, the harder you should lean toward self-building off Amazon's proprietary scaffolding. The more your constraint is speed-to-market with a manageable competitive exposure, the more ASA earns its keep. And for sellers whose catalog lives on someone else's feed, the agent is premature — fix the data first.

"Catalog readiness" is too abstract to act on, so here it is as a scoreable audit. Run each attribute category against your own product feed before you scope any conversational agent. The columns tell you the target completeness, what the agent does when the field is missing, the remediation effort, and how much that category moves conversational discovery.