THE SIGNAL IN ONE SENTENCE
Amazon has turned Seller Assistant from something a merchant asks into something that keeps working after the merchant logs off. The company announced on September 23 that sellers can create continuous workflows in ordinary language, tell them what to monitor and set boundaries around what they may do. One workflow might watch a seller's ten most important products and draft a response when a rating falls below a chosen threshold. Another might scan a category for an opening, propose a price change and refresh a listing. Amazon says each action receives an audit trail, and sellers can choose whether a workflow only recommends or moves toward action. The company also says sellers review and approve actions before they are carried out. That boundary is the heart of the product, not a little safety garnish on top. A price alert is information. Changing a price can alter demand, margin and a competitor's response before breakfast. A restock suggestion is analysis. Sending the wrong inventory into Amazon's fulfillment network can create fees, cash-flow pressure and weeks of cleanup. The assistant now carries what Amazon calls persistent memory of a seller's pricing patterns, inventory cycles and growth goals. That memory follows the seller across Seller Central, Amazon Quick and, through a beta plugin, Anthropic's Claude. Amazon says the plugin can expose listing contributions, real-time performance metrics, inventory levels and sales analytics, while keeping access bounded, requiring human approval for actions and maintaining audit trails. Reuters says the service is optional and free for sellers. Amazon is also offering primary account holders a free year of Quick Plus through December 31, with two additional coworker seats, so the longer-term cost of that particular companion is a separate question. The plain signal is not that a tireless agent finally learned to check a spreadsheet. It is that the marketplace, the seller dashboard, the recommendation engine and the rule-enforcement venue are becoming one continuous operating surface. That can be genuinely useful for a two-person shop. It also means permissions, evidence and rollback have to be designed before the bot gets its first night shift.
01
WHAT ACTUALLY CHANGED
Amazon announced the next version of Seller Assistant on September 23 during its 2026 seller event.
Seller Assistant now supports custom workflows that run continuously, including when the seller is not logged in to Seller Central.
Sellers describe a workflow in natural language and set guardrails around the conditions it should monitor and the response it should prepare.
Amazon's examples include watching product ratings, checking inventory thresholds, monitoring competitive pricing, reviewing account-health alerts and finding product-expansion opportunities.
Prebuilt templates can monitor account health, summarize orders, analyze pricing trends and look for new product opportunities. Sellers can also create workflows from scratch.
A seller can configure a workflow to surface recommendations or move toward actions on the seller's behalf.
Amazon says sellers review and approve actions before they are carried out and that every action receives a complete audit trail.
Seller Assistant now stores persistent context about pricing patterns, inventory cycles and growth goals rather than starting over with each conversation.
Amazon says that context follows a seller across Seller Central, Amazon Quick and Anthropic's Claude.
A new Selling Partner plugin connects listing contributions, performance metrics, inventory levels and sales analytics to Quick and, in beta, Claude.
The plugin is initially in beta for sellers in Amazon's US stores. Amazon says international expansion will follow.
Amazon says the plugin lets a connected assistant act on the seller's business with bounded access, human approval and audit trails.
Amazon says Seller Assistant is built on Amazon Bedrock using Anthropic Claude models and keeps seller data within Amazon's infrastructure.
The company says Seller Assistant capabilities have reached more than 90 percent of selling partners worldwide, have hundreds of thousands of active users and receive recommendation acceptance above 90 percent. Those are company-reported adoption figures.
Reuters reported that the new agent is optional and free for sellers. Amazon separately offers a free 12-month Quick Plus subscription to eligible primary account holders through December 31, 2026.
02
WHY THIS MATTERS
Continuous monitoring is useful because inventory, prices, reviews and account alerts do not wait for a small-business owner to open a laptop.
The move from a chat session to a persistent workflow changes the risk. A bad answer is one event. A bad rule can repeat all night across many products.
Recommendation and execution are different permission levels. The interface should make that distinction visible before a workflow is activated, not bury it inside a settings page.
A human approval step only helps if the seller can see the evidence, affected products, financial range and intended changes without reconstructing the agent's reasoning from scratch.
Natural-language instructions are convenient but ambiguous. A phrase such as keep prices competitive needs a floor, a maximum change, a time window and an exception for promotions or low inventory.
Persistent memory can improve recommendations by preserving business context. It can also preserve stale assumptions, unusual one-time events and preferences the seller no longer wants applied.
Amazon operates the marketplace, controls many of its rules and fees, measures seller performance and now supplies the agent interpreting those signals. That concentration can reduce friction while making independent verification more important.
A pricing workflow may optimize sales while ignoring margin, advertising spend, return rates, fulfillment fees or a seller's cash-flow needs unless those constraints are supplied explicitly.
A rating alert can be valuable, but a falling average does not by itself explain whether the problem is product quality, shipping damage, listing confusion, review manipulation or random variation.
Connecting Seller Assistant to Quick or Claude can bring Amazon data beside accounting, supplier and manufacturing records. It also expands the permission surface across tools and accounts.
Amazon says plugin interactions have clear access boundaries, human approval and full audit trails. The public announcement does not publish a complete field-level access matrix, retention schedule or independent security assessment.
The reported 90 percent recommendation-acceptance figure measures acceptance, not accuracy, profit improvement or seller welfare. People can accept a recommendation because it is convenient, persuasive or difficult to audit.
Small teams may gain operating capacity that once required dedicated analysts. The human consequence should be less repetitive checking, not a new obligation to babysit a faster stream of opaque suggestions.
When the agent makes an error, the seller needs a direct repair path. An audit trail is useful only if it supports rollback, dispute resolution and a clear record of who approved what.
The marketplace also needs to distinguish actions taken by the seller, a coworker, Seller Assistant and an external connected agent so responsibility is not assigned to the wrong person after an incident.
03
WHERE IT COULD HELP
- Begin with alert-only workflows for ratings, account health, stock thresholds and unusual fee changes before enabling any operational action.
- Write every workflow as a policy with a trigger, evidence requirement, allowed products, maximum scope, time window, approval role and stop condition.
- Require a dry-run preview showing which actions the same rule would have proposed during the previous 30 days.
- Set hard price floors from unit cost, fees, advertising spend and required margin rather than telling an agent to stay competitive.
- Use separate approval limits for drafting a listing, publishing it, changing a price, creating a promotion and sending inventory to fulfillment.
- Let the owner approve high-impact changes while staff can approve alerts, drafts and low-risk corrections within defined boundaries.
- Show the source data, timestamp and confidence behind each recommendation, including missing or stale fields.
- Attach an expiration date to persistent assumptions such as seasonal demand, supplier lead time or a temporary growth target.
- Create a visible memory panel where sellers can inspect, correct and delete the preferences Seller Assistant has retained.
- Assign a distinct identity to every workflow and connected agent so audit records do not collapse all automated actions into the seller's primary account.
- Export audit records outside the marketplace so a seller can reconcile them with accounting, inventory and customer-service systems.
- Provide one-click rollback for reversible actions and a documented recovery route for inventory or compliance actions that cannot simply be undone.
- Alert on clusters of related changes, such as a price drop followed by margin compression and stock acceleration, instead of reporting each metric in isolation.
- Measure false alerts, rejected recommendations, corrected memories, prevented errors, time saved and net profit impact rather than only adoption.
- Pause a workflow automatically after repeated rejection, missing data, unexpected scope growth or a material platform-rule change.
- Review connected-tool permissions quarterly and immediately after an employee, agency or contractor loses access to the seller account.
KEEP A HAND ON THE WHEEL
Amazon's September 23 announcement is the primary product account. It says continuous workflows can monitor and respond around the clock, sellers can choose recommendation-only or action-oriented behavior, and actions require seller review and approval. Amazon also says each action has an audit trail, plugin access is bounded and seller data remains protected. The company has not published independent accuracy, false-alert, margin, inventory, retention, security or seller-outcome measurements for the new workflows. Its adoption figures, including reach above 90 percent, hundreds of thousands of active users and recommendation acceptance above 90 percent, are company claims. Acceptance does not establish that a recommendation was correct or profitable. Reuters separately reported that the service is optional and free. Amazon's free 12-month Quick Plus offer has a December 31 signup deadline and does not establish the later subscription cost. Watch for a complete permissions matrix, exact approval behavior for each action type, memory inspection and deletion controls, plugin retention terms, workflow pause and rollback tools, identity records for external agents, international availability, independent seller testing and evidence that the product improves outcomes rather than simply increasing the number of automated decisions.
04
TERMS WORTH KEEPING
OPEN GLOSSARY CARD
Agentic workflow
A continuing automation that watches for conditions, reasons about them and may prepare or perform actions within defined limits.
OPEN GLOSSARY CARD
Persistent memory
Information an AI system retains across sessions so later recommendations can use earlier preferences, patterns or decisions.
OPEN GLOSSARY CARD
AI guardrail
A technical, operational or policy control intended to limit harmful model access, behavior or use within a defined boundary.
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on September 24, 2026.
PUBLICATION RECEIPT: Revision 1. Published September 24, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US