Amazon is expanding Seller Assistant from a system advising sellers into a tool that can remember business context, monitor conditions over time and run ongoing workflows. Meanwhile, the company is launching a Selling Partner plugin connecting seller data to Amazon Quick and entering beta in Claude too.
The update was published on September 23 and shows how Amazon is trying to make an e-commerce assistant into a more agentic system, while retaining layers of seller authorization and oversight.
Persistent business memory
Seller Assistant can now retain business context, such as pricing patterns, inventory cycles and growth goals. This means sellers do not have to start every conversation from scratch or explain the same details each time.
Amazon says this information lets the assistant identify connections over time and suggest actions based on the business's situation. For example, it can track account health, orders, prices or opportunities to expand the catalog.
Processes running in the background
The more significant addition is the ability to set up workflows that continue monitoring conditions even when the seller is not actively conversing with the assistant.
The system can identify a change, present a recommendation or request permission to take an action. Amazon emphasizes authorized actions and an audit trail, allowing sellers to see what was done and in what context.
This is not the same as letting AI manage a store independently. It has not been announced that every action occurs without approval, and not every type of action is necessarily available for automation. The approval level depends on the process and configured permissions.
A new plugin for Claude and Amazon Quick
Amazon also launched a Selling Partner plugin that lets sellers use store data inside Amazon Quick and, in beta, inside Claude.
With authorization, the plugin can expose information about listings, inventory, performance and sales. The idea is that sellers can combine Amazon information with other sources already connected to their workspace, such as accounting or supplier data.
Initially, the Claude beta is intended for sellers in Amazon's US stores. Amazon says additional integrations and international expansion are planned but gave no dates.
An Amazon Quick promotion
According to the announcement, primary account holders can apply for a free Quick Plus subscription for 12 months until December 31, 2026. The research dossier did not specify the price after the promotional period for every user and scenario.
Adoption claims need context
Amazon says Seller Assistant is available to more than 90% of selling partners and that users accept more than 90% of its recommendations. These are company-published figures, not an independent measurement of accuracy, revenue impact or error rates.
No independent figures were published on reversed actions, pricing mistakes, incorrect inventory recommendations or cases where sellers choose to ignore the system either.
What changes in practice
The update brings Seller Assistant closer to a model where AI does not wait for a question, but monitors the business and raises actions as they become relevant.
This can save manual work, but also increases the importance of permissions, documentation and boundaries. In a commercial system, a wrong decision about a price, inventory or listing can directly affect revenue.
The important part is therefore not just how many actions the agent can perform, but how much control remains with the seller, which actions require approval and how to review afterward what the system did.
Automation also increases the need for oversight
As the assistant retains more context and operates over time, an error can also persist beyond a single conversation. An audit trail and the separation between a recommendation and an action are therefore not merely interface details, but a central part of how such a system should work.
Sellers will need to understand which information is retained, who in the organization can approve actions and how to reverse an unwanted change. Amazon did not publish error or reversal data in the announcement, so it is not yet possible to independently assess how reliable long workflows will be in everyday work.