Amazon Adds Always-On Workflows to Seller Assistant, and a Claude and Quick Plugin That Hands Seller Operations to AI Agents
Amazon has given Seller Assistant always-on workflows, persistent seller memory and a Selling Partner plugin for Amazon Quick and Claude. We explain what gets automated, where human approval is still required, what is undisclosed, and what it means for marketplace sellers and operators.
Key Takeaways
- On September 23, 2026, Amazon gave Seller Assistant, its AI for third-party sellers, workflows that keep running while sellers are logged out, persistent memory of each seller's business, and a plugin that connects it to Amazon Quick and Claude
- Routine work such as price monitoring, restocking and account health checks moves from an AI that answers when asked to an AI that keeps watch and acts, with sensitive actions such as refunds and buyer messages still requiring human approval
- The tools are free, but workflows and the plugin are limited to Amazon's U.S. stores for now and timing for Japan is undisclosed. Sellers can prepare by starting in recommendation-only mode and writing down their business rules
Seller Assistant Goes From "Acts When Asked" to "Always Watching"
Sellers can now automate routine tasks 24/7 with Seller Assistant and bring their Amazon data and Seller Central insights into the AI tools they already use.
www.aboutamazon.comOn September 23, 2026, Amazon announced an upgrade to Seller Assistant, the AI assistant for third-party sellers on its marketplace. The announcement came from Mary Beth Westmoreland, vice president of worldwide selling partner experience. It has three pillars: always-on "workflows", a memory that keeps learning each seller's business, and a "Selling Partner plugin" that connects to outside AI tools.
Westmoreland told Reuters that the new agent is free and optional, and that sellers can decide how much personal data they share with Amazon.
Seller Assistant arrived in 2023 as a generative AI that answered questions. In September 2025 it gained agentic capabilities that let it take actions on a seller's behalf with permission, and in March 2026 Canvas added dashboards built from natural language. Even so, it remained a tool that moved only when a seller spoke to it. With this change, a seller gives an instruction once and it keeps working after the seller closes the screen.
What Workflows Automate, and Where They Stop
Amazon's examples are concrete. A simple one: "Alert me if any of my top 10 products drop below a 4-star rating and draft a response plan." A more complex one: "Monitor my top category for competitive openings. If you see one, adjust my pricing and refresh the listings." Sellers describe what they need in plain language, set guardrails (conditions the AI must not cross), and the workflow goes to work.
The operating details live on the Seller Assistant Workflows help page in Seller Central. Workflows run on a schedule or on demand, and sellers can list them, check run history, start a run, or archive them from the conversational Seller Assistant. Five prebuilt templates are available.
| Template | Description on the help page | Routine task it replaces |
|---|---|---|
| Account Health Monitor | Tracks the account and flags issues ranked by urgency | Periodic account health checks |
| Order & Returns Digest | Summarizes recent orders and returns and surfaces urgent items such as late-ship risk and repeat-return products | The morning order and returns review |
| Pricing Analyzer | Scans active listings to find where the seller is missing the Featured Offer | Manual competitive price monitoring |
| Product Expansion Finder | Analyzes what the seller already sells to surface new product opportunities | Assortment expansion research |
| Review & Feedback Requestor | Automatically requests reviews and seller feedback on eligible orders | Sending review requests |
Building a custom workflow is just as simple. A seller clicks "Create Workflow" and describes the task in words, and Seller Assistant asks follow-up questions to fill in the details. The seller then checks the output in a preview that uses simulated data and activates the workflow, at which point it runs on the live account.
What matters in practice is where the AI stops. According to Amazon's announcement, sellers choose, for any prebuilt or custom workflow, whether it only surfaces recommendations or takes actions on their behalf. Even when it takes actions, sellers review and approve them before they are carried out. The help page goes further: sensitive actions such as issuing refunds or sending buyer messages always require approval from the seller or a designated secondary user. If a request falls outside the policy the seller set, the workflow stops and asks rather than deciding on its own. Every action is logged in an audit trail.
The other building block is the "knowledge base." When sellers upload business documents such as their return policy, shipping preferences or SKU-specific guidelines, workflows search them at run time and act on that seller's own rules rather than general knowledge. Sellers who want to hand more work to the AI first need to put their unwritten rules on paper.
One point is easy to miss. The help page states that the primary account owner and all secondary users can view, edit, approve and execute every workflow on the account. Sellers who have given secondary-user access to agencies or outsourced operators should audit those permissions first.
A Memory-Equipped Assistant, and How to Read the "90% Acceptance" Figure
The new memory underpins the workflows' judgment. Amazon says Seller Assistant now retains each seller's pricing patterns, inventory cycles and growth goals, and that its recommendations become more personal with every interaction. The memory carries over to Quick and Claude, and every suggestion comes with its reasoning.
It is built on Anthropic's Claude models on Amazon Bedrock, and Amazon says seller data stays within its infrastructure and is never shared outside Amazon. Amazon disclosed that Seller Assistant has reached more than 90% of its selling partners worldwide in their native languages, with hundreds of thousands of active users and a recommendation acceptance rate above 90%.
That acceptance figure deserves caution. Amazon has not said which kinds of recommendations it measured, or how accepted recommendations affected sales or profit. If the mix includes many answers to questions and routine suggestions, a high acceptance rate follows naturally. "Sellers accept 90%" is evidence of use, not evidence of results.
A Socket for Quick and Claude
The most structural change is that Seller Assistant's intelligence now leaves Amazon's own screens. Amazon says 90% of its selling partners already use third-party AI tools, and it has set the goal of letting sellers run their business with the AI agent of their choice.
The Selling Partner plugin connects Amazon data such as listing contributions, real-time performance metrics, inventory levels and sales analytics directly to the seller's AI agent. The key point is that the agent can take actions, not just read data, the same way Seller Assistant does inside Seller Central. Amazon says access boundaries, human approval of actions and audit trails remain in place through the plugin.
There are two initial destinations. In Amazon Quick, AWS's AI for work, sellers can combine Amazon data with finance, marketing and manufacturing data and build their own pricing, listing or restock agents. Quick shows what it intends to do and waits for approval before acting. Every primary account holder worldwide can sign up by December 31, 2026 for a free 12-month Quick Plus subscription, and can extend it to two co-workers.
The Claude connection is in beta, takes roughly 60 seconds and requires no coding. Sellers who already connect accounting software and supplier systems to Claude can add Seller Assistant itself. Eric Burns, field CTO at Anthropic, said the plugin brings Seller Assistant's insights on pricing, inventory and ads into Claude alongside those existing accounting and supplier connections. The plugin as a whole is in beta for sellers in Amazon's U.S. stores, with international expansion "to follow." FashionNetwork described the overhaul as preparing Amazon's offering for the advent of agentic commerce.
Seller Agents Are Invited In, Buyer Agents Are Screened
On the buyer side, Amazon has built its own agents: the Rufus shopping assistant, Alexa for Shopping, which unified Rufus and Alexa in May, and Buy for Me, which finds and buys items on external brands' sites. Amazon stresses that Buy for Me identifies itself and lets brands opt out.
Three days before the announcement, Amazon cut off Meta's personal agent Muse from shopping on Amazon.com. According to GeekWire, Amazon cited Meta's failure to notify it, Muse not identifying itself while browsing, and the agent appearing to capture and store customer credentials, and it now shows users a message citing its Conditions of Use. Set next to the seller announcement, the pattern becomes clear.
Amazon does not object to outside AI operating an Amazon account as such. On the seller side, it now lets Claude, an outside AI, take actions. The line it draws is whether the agent comes through an official entrance Amazon provides, with identity, scope of action and logging under control. Seller agents are invited in under Amazon's management, while unmanaged buyer agents are kept out. The marketplace appears to be holding on to control in the agent era through how it designs its entrances.
Reservations From the Seller Floor
The promoter's account alone would mislead. About two months ago, a Seller Central forum thread titled "Seller Assistant is a huge fail" collected complaints that getting basic help took a long time and ended with a handoff to a human anyway. One reply said the features Amazon had showcased seemed far down the roadmap from what was actually operational.
Amazon itself adds careful caveats. The Seller Assistant help page says AI is experimental and can make mistakes, and that Amazon is not responsible for claims arising from its responses. The announcement does not address who bears responsibility if a workflow cuts a price too far or misjudges a restock. It is a tool to use on the assumption that final responsibility stays with the seller.
What the Announcement Does Not Disclose
- Availability outside the U.S., including Japan: undisclosed (workflows are U.S. store only; the plugin is a U.S. beta with international expansion "to follow")
- Limits on workflow runs or concurrent workflows: undisclosed
- Quick Plus pricing after the free 12 months: undisclosed
- Claude plan requirements: undisclosed
- Names and timing of additional AI tools: undisclosed
- Compensation and liability if the AI executes a wrong price change or restock: undisclosed
- How the 90%+ acceptance rate is calculated, and its effect on sales and profit: undisclosed
What This Means for E-commerce Businesses
For sellers on Amazon's Japan store, workflows are not yet available. Even so, if they write down their unwritten rules, such as return acceptance criteria, price floors and reorder points, those documents can go straight into a knowledge base. Once the feature arrives, the sensible order is to run it in recommendation-only mode first, check the quality of its suggestions, and only then allow it to act.
For marketplace operators, the bar has moved up a notch. Sellers will come to expect tools that monitor and act for them, include approval and logging, and can be operated from outside AI. Trade publication MDM argues that agentic AI is moving rapidly from separate tools into the systems where commercial work actually happens.
Brands whose main channel is their own store are not exempt either. If competitors on Amazon start handing 24-hour price monitoring and restocking to AI, differences in how fast sellers react on price and inventory will become a bigger competitive factor.
Conclusion
With this upgrade, Seller Assistant has widened its role from an advisor inside Amazon's screens to an agent that lives inside the seller's operations. Approvals, audit trails and knowledge bases form the scaffolding for handing more work to the AI step by step.
Three things are worth watching next: when workflows and the plugin reach other countries, including Japan; which outside AI connects after Quick and Claude; and whether Amazon shows results in sales and time saved rather than acceptance rates.
