Amazon on September 23, 2026 gave Seller Assistant, the AI agent built into Seller Central, a persistent memory of each seller's business and a set of automations that keep running after the seller logs off, and connected the same assistant to Amazon Quick and, in beta, to Anthropic's Claude through a plugin for sellers in its US stores.

In Short

Amazon has given the AI helper inside its seller dashboard a memory of each shop, along with the ability to keep watching things like stock and prices and to prepare or make changes even when nobody is logged in. That matters for the businesses selling on Amazon and for the agencies and advertisers working with them, because software built by Amazon is taking a larger part in price, stock, listing and advertising decisions. Amazon says a person still approves actions before they happen, the link to Amazon Quick and Claude is a US-only beta for now, and a free year of Quick Plus is open worldwide to sellers who sign up by December 31, 2026.

What Amazon published

The changes arrived on the second day of Amazon Accelerate 2026, the company's seventh annual seller conference, held from September 22 to 24 in Seattle. The fuller account is a post on Amazon's corporate news site by Mary Beth Westmoreland, vice president of Worldwide Selling Partner Experience; a condensed version went up on the Seller Forums under the News_Amazon account.

The package has four parts: persistent memory of each seller's business; custom automations, branded Seller Assistant workflows, that monitor an account and respond unprompted; a selling partner plugin carrying Seller Assistant into Amazon Quick, the AWS workplace assistant, and into Claude; and a free 12-month Amazon Quick Plus subscription for every primary account holder worldwide, plus two delegated users, for sign-ups by December 31, 2026.

Westmoreland opens with adoption numbers. According to Amazon, Seller Assistant's capabilities have rolled out to more than 90% of its selling partners worldwide in their native language, the tool now has "hundreds of thousands of active users," and sellers accept its recommendations more than 90% of the time. From those figures the post draws its thesis: "Sellers don't just want AI that works with them. They want AI that works for them."

A short history with an unexplained start date

Amazon's post says the company first put Seller Assistant in front of sellers in 2023, as a generative AI companion. PPC Land's record begins a year later, when Amazon presented a generative AI seller assistant codenamed Project Amelia on September 19, 2024, built on Amazon Bedrock. The new post does not mention Project Amelia and does not explain how the 2023 product and the 2024 one relate.

The larger shift came on September 17, 2025, when Amazon turned Seller Assistant into a system able to monitor accounts, manage inventory and handle advertising tasks with seller approval, available at first to sellers in the US. That move took Amazon's seller tools into agentic AI, and PPC Land's coverage at the time described the system as running on Bedrock with both Amazon Nova and Anthropic Claude models. In April 2026, the assistant added a check on review-sharing eligibility for individual product variations.

Usage figures have since become less precise, not more. Amazon's 2025 Small Business Empowerment Report, published on April 30, 2026, credited Seller Assistant with more than 230,000 monthly users in 2025 and a recommendation acceptance rate above 90%. Five months later, the September post swaps the exact count for "hundreds of thousands" and, a few paragraphs on, describes the task as delivering intelligence "to millions of sellers." Neither number comes with a method.

Memory, reasoning and a canvas

The engineering description is brief. According to Amazon, the assistant is built on Amazon Bedrock with Anthropic's Claude models, developed with AWS and Anthropic, and pairs recent foundation-model work with "25 years of Amazon commerce expertise." The 2026 post names Claude alone. Whether Nova models still sit in the stack, as they did a year ago, is not stated.

Memory is the main technical change. Seller Assistant now retains each seller's pricing patterns, inventory cycles and growth goals, and that context travels: it follows a seller across Seller Central, Amazon Quick and Claude, so switching tools does not reset it. Amazon describes recommendations that "sharpen with every interaction," and says every suggestion arrives with its reasoning attached so a seller can evaluate it, adjust it or ask for alternatives.

Amazon's worked example shows how far a single question now reaches. Asked how a product is performing, the assistant can identify a competitive signal, link it to an advertising or cross-promotion opportunity, assess possible improvements to the listing and draft a recommendation. For sellers who prefer to see numbers laid out, there is a canvas, a visual workspace in which scenarios can be modelled, pricing strategies compared and campaign performance diagnosed on live data, all through conversation.

The post is silent on where the memory is stored, how long it persists and whether a seller can inspect, correct or erase it. Its only data commitment is broad: Seller Assistant, according to Amazon, keeps seller data within Amazon's infrastructure and never shares it outside the company.

Automations that do not wait to be asked

Until now, the assistant acted on request. A seller could type "adjust price for this ASIN" or "draft a restock plan," and it would reason through the problem and execute with permission. Amazon's case for the change is that recurring checks - inventory thresholds, competitive pricing, account health alerts - follow consistent patterns and are repetitive enough to run on their own.

The new automations run continuously, monitoring conditions, interpreting changes and responding, "even when the seller isn't logged in to Seller Central," according to Amazon. A seller describes the job in plain language and sets guardrails. Amazon gives two examples at opposite ends of the range. The modest one reads: "Alert me if any of my top 10 products drop below a 4-star rating and draft a response plan." The ambitious one goes further: "Monitor my top category for competitive openings. If you see one, adjust my pricing and refresh the listings."

Every action is logged with a complete audit trail. Prebuilt templates cover four jobs - monitoring account health, summarising orders, analysing pricing trends and finding product expansion opportunities - and custom automations can be built from scratch.

Recommend, act, or wait for sign-off?

For each automation, prebuilt or custom, the seller chooses whether it only surfaces recommendations or takes actions. Here the two Amazon texts diverge. The forum summary presents a clean choice between recommending and acting, with every action logged. The longer post attaches a condition to the second mode: when an automation takes actions, "sellers stay in control, reviewing and approving them before they're carried out."

Read literally, that turns the acting mode into a queue for human sign-off rather than unattended execution. It sits awkwardly beside the promise of software that responds while the seller is away, and beside the second example, in which a detected opening leads straight to a price change and a listing refresh. The post does not say whether approval is given per action or once at set-up, where approval requests are delivered, how long a pending action waits, or what happens to a time-sensitive price move nobody approves. Nor does it list what a guardrail can express, whether a price floor, a cap on daily changes or a spending limit.

Amazon has drawn this line before: with Ads Agent, unveiled at unBoxed in November 2025, campaigns only went live after advertisers reviewed and approved its recommendations.

Not every reader took the approval language at face value. Jon Elder, who describes himself on LinkedIn as an Amazon and Walmart advisor for elite brands and a three-time founder, and whose commentary PPC Land cited during the April 2026 fuel surcharge dispute, posted under the heading "Amazon AGI is here." His verdict: "It's both terrifying and exciting!" Elder treated unattended action as a setting sellers could switch on, warned against letting Seller Central act without the seller's permission - "This is a recipe for disaster" - and urged sellers to "Keep everything as 'recommend-only.'" He added that "AI cannot be trusted with full owner control." He also shared seven prompts, starting with one on funnel ratios in the Search Query Performance report. Amazon's published text describes no mode that acts without prior approval, and Elder's post does not point to one.

A plugin for outside assistants

The plugin rests on one statistic. According to Amazon, 90% of its selling partners already use third-party AI tools to run their businesses; the post does not say how that was measured. Amazon's stated aim is that sellers can use the AI agent of their choice.

The plugin links listing contributions, real-time performance metrics, inventory levels and sales analytics to that agent, and it is not read-only. "Their AI doesn't just read from their Amazon business; it can act on it, the same way Seller Assistant does inside Seller Central," the post says. Each plugin interaction, according to Amazon, carries limits on what it can access, human approval on actions and a complete audit trail. It is available in beta to sellers in Amazon's US stores, with international expansion and further integrations promised but not dated. Even the name wavers: the post calls it the selling partner plugin and the Amazon Selling Partner plugin, while Anthropic's quoted executive calls it the Seller Assistant plugin for Claude.

Advertising data is absent from the list of what the plugin connects. Amazon Ads runs its own route for assistants: its MCP server entered open beta on February 2, 2026, connecting Claude, ChatGPT, Gemini, Amazon Q, Bedrock and AgentCore to advertising functions through the Model Context Protocol. The seller-side plugin starts narrower, with two hosts, one of them Amazon's own, and Amazon does not say which protocol it uses.

Amazon Quick and the $720 figure

Quick, according to Amazon, writes documents and spreadsheets, runs deep research and connects to email, calendars and files. With the plugin, sellers can set Amazon data beside financial, marketing and manufacturing records and build pricing, listing or restock agents inside Quick, which shows what it intends to do and waits for approval before acting. Jigar Thakkar, vice president of Agentic AI for Business at AWS, said the plugin keeps "continuously learning their preferences to give them answers and take actions that reflect the full picture of their business."

Quick has been widening its reach through 2026. AWS added a desktop application with a persistent personal knowledge graph on April 28, in a post written by Thakkar, and in July Amazon added Quick to its Business Prime benefits.

The sign-up page puts a price on the seller offer: "1 year free up to $720 value," with a footnote specifying "A full year free for primary sellers and 2 collaborators." The arithmetic follows AWS's list prices. According to AWS, Plus costs $20 per user per month on annual billing, or $240 a year, and $25 per month on monthly billing; three annual seats come to $720. Claiming it means accepting the AWS Customer Agreement, a contract separate from the one governing Seller Central. None of Amazon's materials says whether the subscription renews at list price after 12 months or simply lapses.

Geography cuts the other way. The Quick Plus offer is global, while the plugin that feeds Quick its Amazon data is limited to US stores, so a seller in Germany or Japan can take the free year now without the connection the offer is built around.

Claude, in beta

Many sellers, according to Amazon, already run parts of their business through Claude, linking channel data, accounting software and supplier systems. With the plugin, those sellers get Seller Assistant itself inside Claude, connected in roughly 60 seconds without coding. Eric Burns, field CTO at Anthropic, restated the adoption claim in his own words: "Amazon Seller Assistant already uses Claude models to help sellers make informed, strategic decisions, and they accept its recommendations more than 90% of the time."

The post's only customer voice is a beta tester chosen by Amazon: "What took an afternoon now happens instantly," said James Oliver, an Amazon selling partner and founder of Atlas Bar. No measured time savings are published.

The commercial ties behind the choice of partner are substantial. Amazon completed a $4 billion investment in Anthropic by March 2024, and after March 31, 2026 it put a further $5 billion into Anthropic preferred stock.

Where the data ends up

The promise that Seller Assistant keeps data inside Amazon's infrastructure sits in the same post as a product built to move that data into an outside assistant. Amazon's answer is that external connections "don't change how Amazon protects seller data." The post does not address what happens to listing, inventory and sales data once it sits in a Claude account - retention, training use, visibility to other connectors - all matters governed by Anthropic's terms. PPC Land reported in June that Anthropic's updated privacy policy covers inputs and outputs Claude sends to third-party services through connectors and plugins.

Software Amazon did not build faces stricter terms. An Agent Policy in the Business Solutions Agreement, effective March 4, 2026, requires AI agents accessing Amazon services to identify themselves as automated systems, and on August 4 the Ninth Circuit vacated the injunction Amazon had won against Perplexity's Comet shopping agent. Independent developers who connect seller accounts to outside tools were told in November 2025 to expect a $1,400 annual Selling Partner API fee, which Amazon dropped on May 14, 2026 after developer pushback. How the plugin will coexist with the 2,500-plus applications built on that API, the post does not say.

Loose ends in Amazon's own materials

Beyond the start date, two details do not line up. The post's summary box lists Quick as the first host and Claude as the beta one, yet its body puts the plugin as a whole in beta for US stores, and the forum summary marks both as beta. And three separate 90% figures - rollout coverage, recommendation acceptance and third-party AI use - measure different things, none with a published method. Pricing after the beta is absent altogether: no Amazon text says whether the automations or the plugin will carry fees once testing ends, the very question one forum reply raised.

Reaction on the Seller Forums

At the time the forum thread was captured, the post had 137 views, four replies, one upvote and ten downvotes. One reply was collapsed; no response from Amazon was visible.

The replies were not warm. One seller predicted that sellers would debug the tools for free before Amazon charged for them, and that listings built with a paid Amazon service would get an edge: "I can't wait!!" Another asked whether the tools would "Actually be helpful?" and "Stop customers from scamming us?", ending a list of six on rising fees. A third reported that "the sign up link doesn't work."

The fee complaint has a recent reference point. A 3.5% fuel and logistics surcharge has applied to Fulfillment by Amazonfees since April 17, 2026, averaging about $0.17 per US unit, and it has drawn objections in later forum threads as well.

Why this matters for the marketing community

Advertising sits at the centre of Amazon's own example, in which a question about product performance ends with an advertising or cross-promotion opportunity. That assistant is built by the company that sells the ads, whose advertising revenue reached $19.8 billion in the second quarter of 2026, up 26%, and by Amazon's count sellers accept its recommendations more than 90% of the time. For agencies and in-house teams managing Sponsored Products budgets, a persistent assistant with standing automations inside the seller's own console becomes another participant in how that money is allocated, alongside Amazon's Ads Agent on the advertiser side.

Timing adds a second angle. The automations arrive two weeks before Prime Big Deal Days, which Amazon set for October 6 and 7 across 22 countries, with Alexa for Shopping able to watch prices and buy once a product drops to a chosen level. Amazon-built software will therefore sit on both sides of the same price: one agent watching it for shoppers, another set up to move it for sellers. How will deal eligibility, reference prices and bids behave when both ends are automated? The post does not address event periods at all.

Listings are the third. The ambitious example ends with "refresh the listings," and listing text feeds the relevance that search ads depend on. Amazon's recent precedent is instructive: from July 27, 2026, its own AI rewrote product titles above a new 75-character limit, with a 14-day window for brand owners to review the changes. The automations move that kind of editing under rules the seller writes.

Last comes consolidation. Meta opened its ad system to Claude and ChatGPT on April 29, 2026 and Semrush published a native Claude connector on August 26, so an agency working in Claude can now hold Meta campaign data, Semrush keyword data and Amazon selling data in one conversation, with Amazon ads reachable separately through the Ads MCP server - but the first two doors to Amazon's seller data open onto Amazon's own assistant and a model maker it has funded. More integrations will follow, Amazon says. It has not said which, or when.

Timeline

Summary

Who: Amazon, through Mary Beth Westmoreland, vice president of Worldwide Selling Partner Experience, with AWS (Jigar Thakkar, vice president of Agentic AI for Business) and Anthropic (Eric Burns, field CTO). Affected parties are Amazon's third-party sellers, the agencies and advertisers working with them, and the developers who build seller tools.

What: Seller Assistant gains persistent memory of pricing patterns, inventory cycles and growth goals; continuous automations with prebuilt templates, a recommend-or-act choice, seller approval and audit trails; and a selling partner plugin, in beta, that links listing, performance, inventory and sales data to Amazon Quick and Claude. Primary account holders worldwide can claim 12 months of Quick Plus for three seats, valued by Amazon at up to $720.

When: Published on September 23, 2026, during Amazon Accelerate 2026 (September 22-24). The Quick Plus sign-up window closes on December 31, 2026.

Where: The plugin is available in beta to sellers in Amazon's US stores, with international expansion undated. The automations sit inside the new Seller Central experience, and the Quick Plus offer applies globally.

Why: Amazon says 90% of its selling partners already use third-party AI tools and that sellers want AI that acts for them rather than with them. The move extends Amazon's own assistant into outside tools on Amazon's terms, while leaving open questions on approval mechanics, data handling outside Amazon, post-beta pricing and the influence of an Amazon-built assistant over advertising and pricing decisions.