TLDR;
AWS has added seven capabilities to its Agentic Shopping Assistant, including customer memory, image and voice search, cart controls, Shopify connectivity and operational dashboards. It gives retailers a way to keep conversational shopping inside their own site and AWS account instead of handing the customer relationship to a general-purpose assistant. The tradeoff is that brands now own the difficult work around consent, product data, guardrails and proving that the assistant improves purchases.
What happened
AWS says it has shipped seven upgrades to the Agentic Shopping Assistant on AWS. The package now includes a self-service admin portal, long-term memory, image and voice search, reusable conversation history, cart management, Shopify connectivity through the Universal Commerce Protocol and conversational intelligence dashboards. This moves the product beyond product recommendations and closer to a complete journey from a vague customer need to a populated basket and checkout.
The control model is aimed at enterprise retailers. AWS says the code runs in the retailer's own AWS account and customer data is not shared by the solution. Business teams can change brand prompts, ingest catalogue updates, connect MCP tools, monitor tool health and restore earlier versions. The AWS Marketplace listing offers an AWS build, an advisory-led build or partner support, which confirms this is an implementation programme rather than a feature switch.
The most consequential feature is memory. Amazon Bedrock AgentCore can store preferences from conversations, while the assistant can combine them with profile attributes such as loyalty tier, sizes, categories and usual price range. AWS also says its cart layer works with existing commerce infrastructure for logged-in and guest shoppers. Retailers still need to test that against their identity, consent and order systems.
Why it matters
The strategic point is control. Product discovery is moving into ChatGPT, Google, Alexa and other assistants, where the platform decides which products appear and how the journey is measured. AWS offers retailers similar conversational behaviour on their own property, with first-party intent feeding merchandising and marketing. AWS claims the assistant can lift conversion and reduce returns, but this announcement provides no independent evidence for those outcomes.
Owning the experience also means owning its failures. A remembered size can become stale, a product card can show an unavailable variant, and a cart can miss a promotion. Memory helps when it is accurate, but it needs clear consent, deletion controls, identity separation and retention rules. Because AWS organises memory by actor, session and namespace, brands must map those identifiers to customer-data policy.
There is a commercial tension too. Amazon has blocked Meta's Muse from shopping its marketplace while AWS gives retailers tools to build assistants on their own sites. Agent access, customer data and transaction control will remain contested. A first-party assistant is valuable only if customers use it and it does more than put site search in a chat window.
How your brand can benefit / be affected
Start with one category where customers need guidance and product data is reliable. Furniture, electronics, beauty and complex apparel are sensible candidates because choices depend on attributes that standard filters handle badly. Define a narrow job such as finding a compatible product or completing a routine bundle. Do not begin with the whole catalogue.
Treat catalogue, pricing, stock and cart APIs as production dependencies. The assistant should retrieve current facts from authoritative systems, flag missing information and require confirmation before changing quantities or placing an order. Give each MCP tool a limited purpose and log every action. Prompt version history still needs access control, test cases and a rollback plan.
Set the customer-data contract before enabling memory. Decide which preferences can be stored, how they link across sessions, when they expire and how customers can remove them. Separate helpful preferences such as size or budget from sensitive inference. Test that guest, logged-in and shared-device journeys cannot leak one person's history into another session.
Measure the pilot as a commerce product. Track successful product matches, add-to-cart rate, checkout completion, returns, cancellations, latency, tool errors, guardrail triggers and cost per completed task. Use a holdout or controlled rollout so conversion claims have a credible baseline. AWS has supplied more of the stack, but the brand must prove customers prefer it and that the economics beat simpler search, filters and human support.
News date: 26 September 2026. Editorial review: 28 September 2026. Analysis includes subsequent developments where stated.