Amazon's Seller Agent Can Manage Listings While Merchants Are Offline

Amazon is giving marketplace sellers an agent that can watch inventory, pricing, ratings and account health while they are offline. The company says sellers will choose whether the workflow only recommends an action or carries it out after approval.

The change is bigger than a chat assistant. On September 23, Amazon added persistent memory to Seller Assistant, custom workflows that run continuously, and a selling-partner plugin that launches in Amazon Quick and is in beta with Anthropic’s Claude.

For a seller, the marketing implication is concrete. Product listings, availability, price and ad performance are connected decisions. A low inventory level can change a promotion. A drop in rating can change a product-page message. A competitor’s price movement can affect both margin and paid media. The agent is designed to watch those conditions together.

Amazon says its assistant can now remember a seller’s pricing patterns, inventory cycles and growth goals across Seller Central, Quick and Claude. It also says sellers can create a workflow in plain language, set guardrails and get a full audit trail.

The company offers a simple example: alert a seller when one of the top 10 products falls below a four-star rating and draft a response plan. A more consequential version can monitor a category, change pricing and refresh listings. Amazon says the seller approves actions before they are carried out.

That approval sentence matters more than the word “agent.” Amazon says recommendations from Seller Assistant are accepted more than 90% of the time and that hundreds of thousands of sellers actively use it. Those are Amazon’s reported figures. The company hasn’t published the sample, time period, adoption denominator or whether acceptance led to better sales, margin or customer experience.

The new plugin brings the assistant’s listing contributions, performance metrics, inventory and sales analytics into Quick and Claude. Amazon says seller data stays inside Amazon’s infrastructure and that the connections include permission boundaries and audit trails. The initial availability is in Quick and a Claude beta, so a seller should check access before rebuilding a workflow around it.

This is an operations product as much as a marketing product. A marketplace marketer should decide where automation is safe: monitoring stale listings, flagging poor reviews and preparing a restock plan are different from altering prices or publishing product copy.

Start with recommendation mode and a narrow metric. Measure the time from a marketplace signal to a completed correction, then check whether the correction protected revenue or just created extra approval work. Amazon can automate the watch. The merchant still owns the offer.

Start with the work that can be checked

Amazon’s September product update describes a seller-defined workflow that can watch a condition, draft a response and preserve an audit trail. That makes it useful for operational work with a visible before-and-after state. A listing either carried an outdated detail or it didn’t. An inventory alert either reached the owner in time or it didn’t.

The earlier Seller Assistant announcement positioned the product as an agent that could reason, plan and act with permission. The new release gives that claim a more concrete interface: memory across tools, ongoing workflows and a plugin in Quick and Claude.

Amazon’s selling-partner overview says the assistant can help across inventory, advertising strategy and compliance. Those are adjacent but different risk levels. A recommendation to investigate a review is low-risk. A price change, a listing rewrite or a change to an advertising plan can alter revenue, margin and brand positioning. Treat each permission as its own policy.

This is also an access decision. Sellers that build their catalog operations around Amazon’s agent should document which data it can read, who approves an action, and what happens when the workflow is wrong. That matters as Amazon and Shopify take different approaches to outside shopping agents and as AI marketing tools move more campaign work into automated systems.

A practical first test: choose one repeated task that currently produces a ticket or spreadsheet. Let Seller Assistant watch it in recommendation mode for two weeks. Compare the number of problems found, the time to correction, false alerts and the hours spent reviewing suggestions. Only then allow an action that changes a listing, price or campaign.

Seller Assistant makes marketplace operations more continuous. The seller’s controls have to become more continuous too.