Amazon's message to its marketplace sellers on September 23, 2026 was: keep working in whichever AI assistant you like, and we will plug in. At Accelerate, Amazon announced upgrades to Seller Assistant and a plugin that lets sellers run parts of their business from Anthropic's Claude or Amazon's own Quick assistant.
The headline that spread, "Amazon moves AI agents into direct operations for 90% of sellers," is misleading. The 90% numbers in Amazon's announcement mean something else. This guide gives the accurate version, explains what the plugin and workflows do, and lists the guardrails to set before you switch anything on.
TL;DR
| Question | Answer |
|---|---|
| What is new? | Seller Assistant gets persistent memory, 24/7 workflows, a canvas, and a plugin for Claude (beta) and Amazon Quick |
| When? | Announced September 23, 2026 at Amazon Accelerate |
| What can it do? | Monitor conditions, recommend and, within guardrails, execute tasks like price and listing updates |
| Control? | Approval gates, customizable guardrails, audit trails |
| The 90% claim? | Reach (90%+ of sellers), acceptance rate (90%+ of recommendations), and outside AI use (about 90%) are three different numbers |
| Data? | Amazon says seller data stays in Amazon infrastructure |
| Claude status | Plugin is in beta; Quick plugin is launched |
What Seller Assistant can do now
According to Amazon's announcement and coverage from GeekWire, PYMNTS and others:
Persistent memory. The assistant retains information about each seller's pricing patterns, inventory cycles and goals, and that context follows the seller across Seller Central, Amazon Quick and Claude. Recommendations are meant to improve over time.
24/7 workflows. Seller Assistant can run continuously, monitoring conditions such as product ratings or competitor pricing and executing recurring tasks without waiting for a new request, within seller-defined parameters. Every action carries a complete audit trail, and sellers keep control through approval gates and guardrails. Reporting stresses that Amazon still requires explicit approval for actions.
Canvas. A visual workspace generated on the fly from a seller's live data. Sellers can compare pricing strategies or diagnose ad campaigns, and the canvas redraws as they ask follow-up questions.
The plugin. A selling partner plugin connects Amazon data, including listings, performance metrics, inventory and sales analytics, to external AI agents. It is available for Amazon Quick (launched) and Claude (beta). The idea is to let a seller check inventory, adjust prices and update listings without opening Seller Central.
Fact check: what "90%" actually refers to
Three different statistics are in circulation.
| Claim | What it means | Source |
|---|---|---|
| Seller Assistant has reached over 90% of selling partners | Availability across marketplaces, in native languages | Amazon announcement |
| Sellers accept over 90% of its recommendations | When the assistant recommends something, sellers usually approve it | Amazon announcement |
| About 90% of sellers already use some outside AI | Sellers use AI tools from other vendors for parts of their work | Reporting on Amazon's remarks |
None says that agents autonomously run 90% of seller operations. The headline that did was a compression of the three. It matters because the real product still puts a human approval gate in front of actions, and the acceptance rate measures approvals of suggestions, not autonomy.
A caution on the acceptance figure: a 90% acceptance rate can mean the recommendations are good, or that sellers are rubber-stamping. Ask for outcome data, not just approvals.
Why Amazon is opening a plugin to Claude
Amazon has fought outside shopping agents; it blocked Meta's Muse shopping agent. The seller plugin is the opposite posture: a first-party interface into Claude and Quick, with Amazon controlling scopes, approvals and logging.
Strategic reading, which is analysis rather than Amazon's stated reason:
- Sellers already use outside AI tools. Meeting them there keeps Amazon's data and workflow at the center.
- A plugin lets Amazon define what agents can do, unlike scraping or browser automation.
- Partnering with a leading assistant provides distribution without building every interface.
For Anthropic, it is another sign of the Claude ecosystem growing: see the Claude Marketplace expansion and how agent skills work in what agent skills are.
Risks sellers should think about
Wrong price changes at scale. An agent that reprices 5,000 SKUs on a bad signal can do damage quickly. Use price floors and ceilings.
Compliance. Some actions, such as content in listings or claims about products, carry policy risk. Keep approvals for anything customer-facing.
Data exposure. The plugin links account data to an outside agent. Amazon says data is not shared externally, but review permissions and terms, and limit scopes.
Prompt injection. Agents that read reviews, messages or supplier emails can be manipulated by hostile text. See indirect prompt injection.
Overreliance. A 90% acceptance rate can lull teams into approving without reading.
A safe rollout plan
- Start read-only. Use the plugin for reports and analysis first.
- Pilot on a small catalog. Pick 20 to 50 products with stable demand.
- Set hard limits. Price floors, maximum daily change percentages, and blocked categories.
- Require approval for writes. Keep humans in the loop for price, listing and ad budget changes.
- Review audit logs weekly. Look for patterns: repeated overrides suggest bad recommendations.
- Track outcomes. Margin, sell-through, stockouts and ad efficiency, not just time saved.
- Expand gradually. Only widen scope after you have a month of clean results.
What this means for builders
If you build agents or plugins for commerce platforms, Amazon's design is a template:
- Persistent memory across surfaces so context follows the user.
- Approval gates and audit trails as first-class features.
- A plugin interface that exposes data and actions in a controlled way.
- A visual canvas for analysis that a chat transcript cannot provide.
Related infrastructure for agent purchases is developing quickly: Cloudflare Wallets for agent payments, Meta Muse's transaction layer and Shopify checkout with Meta.
Example workflows, and how to keep each one safe
These are illustrative patterns based on the capabilities Amazon describes: monitoring ratings, adjusting prices against competitors, checking inventory and updating listings.
1. Low-stock alert with a reorder draft. The workflow watches inventory against sales velocity and, when days of cover drops below a threshold, drafts a reorder quantity for approval. Safe design: the agent proposes and you approve; it never places the order itself.
2. Competitive repricing inside a band. The agent tracks competitor prices for a defined set of ASINs and proposes changes within a floor and ceiling you set. Safe design: a hard minimum margin, a maximum change per day, and an alert when the agent hits a boundary repeatedly, since that usually means the band is wrong.
3. Rating watcher. The agent flags products whose average rating drops, summarizes the recent negative reviews and suggests fixes such as a listing update or a supplier check. Safe design: read-only, with the summary reviewed by a human before anyone answers customers.
4. Ad diagnosis in the canvas. The canvas compares campaigns and redraws as you ask follow-ups. Safe design: use it for analysis first; move budget only after you confirm the numbers against Seller Central.
5. Listing hygiene. The agent finds missing attributes or inconsistent titles and proposes edits. Safe design: batch changes into a review queue, and sample-check before approving in bulk.
Questions to ask Amazon and Anthropic before adopting
- What exactly can the plugin read and write? Ask for the full scope list.
- Where are prompts, tool calls and results stored, and for how long, on the Claude side?
- Can I revoke access instantly, and does revoking stop scheduled workflows?
- Is there an audit export I can send to my accountant or compliance team?
- Are there rate limits or fees that apply to plugin use?
- How are policy violations handled if an agent action triggers an account issue: who is responsible?
For businesses that sell across several channels, remember that agent plugins are spreading everywhere. If you already use assistants for Shopify or ad platforms, decide on a common policy for approvals, scopes and logging rather than configuring each tool differently.
What we do not know
- Pricing. Whether the plugin or workflows cost extra.
- Claude availability. The plugin is in beta; regions and eligibility are unclear.
- Outcome data. No published evidence that recommendations improve profit.
- Error handling. How the system behaves when signals conflict or data is stale.
Bottom line
Amazon is turning Seller Assistant from a chatbot into an operator with guardrails, and letting sellers reach it from Claude. The claim that agents now run 90% of seller operations is wrong; the accurate version is more interesting: a controlled, auditable agent interface that Amazon owns. Sellers who start read-only and set hard limits can benefit without handing over the store.
Details reflect Amazon's September 23, 2026 announcement and press coverage. The plugin for Claude is in beta and features may change.
Related reading
- Amazon blocks Meta Muse shopping agent
- Meta Muse transaction layer with PayPal, Shopify and Expedia
- Cloudflare Wallets for AI agent payments
- Anthropic Claude Marketplace expansion
- Shopify and Meta agentic Shop Pay checkout
- What are agent skills?
- Meta Muse trust week
- Official: Amazon announcement, GeekWire report