On 23 September, at its Accelerate seller conference, Amazon changed what its Seller Assistant is. Until now it was a chat box inside Seller Central that answered when you asked. It now remembers your business, runs standing instructions around the clock while you are logged out, and plugs into Anthropic's Claude and Amazon's own Quick assistant, so you can check stock and change prices without opening Seller Central at all. That is a real new capability, and for a small seller it is mostly good news. It also moves one decision from "someday" to "this week": which things you let an AI change in your store, and which it may only suggest. This article sets out what Amazon shipped, what you can use today if you sell outside the US, and where we would draw that line, having run always-on AI in production ourselves.
What Amazon Announced at Accelerate on 23 September
Amazon's own announcement lists three changes to Seller Assistant, plus a free offer that matters more than it looks.
- Memory. Seller Assistant now keeps a memory of each seller's "pricing patterns, inventory cycles, and growth goals", and that memory follows you across Seller Central, Amazon Quick and Claude instead of resetting every session.
- Workflows. Custom automations that, in Amazon's words, "run continuously, monitoring conditions, interpreting changes, and responding dynamically, even when the seller isn't logged in to Seller Central".
- The selling partner plugin. It connects your listings, real-time performance metrics, inventory levels and sales analytics to Amazon Quick, and in beta to Claude. GeekWire and SiliconANGLE report that connecting Claude takes about 60 seconds with no coding, and that the plugin can check inventory, adjust prices and update listings, with the seller approving each action before it runs.
- A canvas. A visual workspace inside Seller Assistant for comparing pricing strategies and diagnosing ad campaigns.
- A free year of Amazon Quick Plus. Every primary account holder, worldwide, can sign up for 12 months of Quick Plus at no charge until 31 December 2026, and name two co-workers who get a free year too.
Mary Beth Westmoreland, Amazon's vice president of Worldwide Selling Partner Experience, put the intent plainly to GeekWire: "Our vision was that they would never have to log into Seller Central." That is the headline. The seller dashboard stops being the place you go to run the business, and becomes a thing an assistant reads on your behalf.
The Real Change Is That It Keeps Working After You Log Out
A chat assistant that answers questions is a convenience. An assistant that watches your account at 3 a.m. and acts on what it sees is closer to a member of staff. Amazon's own examples show the difference:
- A watching brief. "Alert me if any of my top 10 products drop below a 4-star rating and draft a response plan." Nothing changes in your store. You get a warning and a draft.
- A standing order. "Monitor my top category for competitive openings. If you see one, adjust my pricing and refresh the listings." This one changes your prices and your listings, on a trigger the AI interprets, at whatever hour the opening appears.
Both are written in the same plain sentence. Both take the same ten seconds to set up. One is a smoke alarm and the other puts your price list in play. Amazon has built the right controls for this: it says "every action is logged with a complete audit trail", that sellers "choose whether it just surfaces recommendations or takes actions on their behalf", and that when a workflow takes actions, sellers stay in control, "reviewing and approving them before they're carried out". The controls exist. Whether you use them well is now your job, not Amazon's.
Amazon also says Seller Assistant has reached over 90% of its selling partners worldwide in their own language, with hundreds of thousands of active users, and that sellers accept its recommendations over 90% of the time. Those are Amazon's figures, not an independent measurement, but the second one is worth sitting with. When a person accepts nine suggestions out of ten, they are no longer reviewing them. They are clicking yes. That is exactly the moment an approval step turns into a formality, and exactly the moment the one bad suggestion in ten stops being caught.
Let an always-on agent tell you anything it likes. Let it change only what you could undo in one click, and would not lose money on if it got it wrong overnight.
Recommend or Act: The One Setting That Decides Your Risk
We run AI that works unattended for our own customers. Our AI Cam watches cameras around the clock, judges what it sees and writes up every event, and every one of those events is logged. Our Accounts module can read a chart of accounts from a photo, and writes nothing until a person approves the preview. Building those taught us three things that apply directly to an Amazon store.
- The AI is usually right. The trigger is what goes wrong. "A competitive opening" is a judgement. A competitor's price glitch, a listing that is briefly out of stock, a rival clearing old stock at a loss: each can look like an opening to a model reading numbers at speed. The model does not have to be bad to act on a bad signal.
- Approvals that arrive at night get rushed. Amazon says actions are approved before they run. But an approval request at 2 a.m. either waits until morning, by which time the opening has gone, or gets tapped through half-asleep on a phone. Decide in advance which requests you will approve on the spot, and set limits that assume you will not read them carefully.
- An audit trail only helps if someone reads it. Amazon logs every action. Put ten minutes in the diary each morning to read what the assistant did overnight, at least for the first month. That is how you find out whether your instructions mean what you thought they meant.
What We Would Let It Change, and What We Would Not
This is our view, not Amazon's guidance, and it assumes a small seller with no one watching the account overnight. The test in the right-hand column is the one we use for our own systems: how bad is a wrong answer, and how quickly can you reverse it?
| Task | Our setting | Why |
|---|---|---|
| Rating drops, review spikes, suppressed listings, stock running low | Act freely: alert and draft | Nothing in the store changes. The worst outcome is an unnecessary message. |
| Listing copy, bullet points, images | Recommend, you approve | Reversible, but a bad claim in a listing can breach Amazon policy or mislead a buyer. |
| Price changes inside a band you set | Act, with a floor and a ceiling | Only once you have read a few weeks of its suggestions and agree with them. |
| Price changes with no floor | Never | One misread signal, approved in a hurry, sells stock at a loss. |
| Ad budgets and bids | Recommend only, to start | Money leaves the account immediately and does not come back. |
| Account settings | Always you | Rarely urgent, and hard to diagnose when changed without you. |
If you do let it set prices, write the floor into the instruction itself, for example "never below landed cost plus 15%", and check in the audit trail that it respected it. A limit you assume is there is not a limit.
The Stock It Can See Is Not All of Your Stock
The plugin reads the inventory Amazon knows about. For a seller who only sells on Amazon, that is the whole business. For most of the businesses we work with, it is one channel among several: a shop counter, a website, WhatsApp orders, a second marketplace, a wholesale customer. An assistant that "refreshes listings" and sets prices from Amazon's view of your stock is working from part of the picture.
- Keep one source of truth for stock. Whatever system holds your real quantities, across every warehouse and channel, is the one the numbers should come from. If that is a spreadsheet, the AI is only as current as the last time someone updated it. A proper inventory system with low-stock alerts across warehouses closes that gap. We covered the channel problem in more depth in our e-commerce inventory guide.
- Margin lives outside Amazon too. Amazon sees your selling price and its own fees. It does not see what you paid your supplier last month or what the shipment cost to land. Any pricing rule you give the assistant should be built on your own cost figures, not on Amazon's.
- The money still has to reach your books. Whatever the assistant sells, the payouts still have to be matched to your bank and your ledger. That part of the job has not changed.
If You Sell Outside the US, Here Is What You Can Use Today
Amazon is clear that the plugin is "available in beta for sellers in Amazon's U.S. stores, with international expansion to follow". It gives no date. So if you sell on Amazon.ae or Amazon.sa, or you are a Pakistani seller listing on the US store through your own account, the position is:
- The free Quick Plus year is global. Amazon says every primary account holder worldwide can sign up, until 31 December 2026. Quick Plus lists at $20 per user per month on an annual plan, or $25 month to month, on Amazon's own pricing page. Three seats for a year is about $720 of list price at the annual rate. There is no reason not to claim it before the deadline, even if you only use it to learn the tool.
- The Claude plugin is US-store only, for now. If your account sells in the US store, you can try the beta. If it sells elsewhere, you wait for the expansion Amazon has promised.
- Seller Assistant itself is already widely available. Amazon says it is live for over 90% of selling partners worldwide. Check Seller Central for which of the new workflow features have reached your marketplace before planning around them.
Our honest read: for most sellers outside the US, the practical gain this month is the free Quick Plus year and the chance to set up watch-only workflows. The Claude connection will matter more once it reaches your marketplace, and you will be better placed to use it if you have already learned how the assistant thinks in recommend-only mode.
What to Do This Week
You do not need to wait for anything to start. In order:
- Claim the free Quick Plus year for yourself and two colleagues before 31 December. It costs nothing and does not commit you to anything.
- Set up two or three watch-only workflows first. Ratings, stock running low and suppressed listings are good starters. Nothing in your store changes, and you learn how the assistant reads your account.
- Keep every workflow on "recommend" for a month. Note how often you would have overruled it. If the answer is rarely, and never on price, move one narrow task to "act", with a floor written into the instruction.
- Read the audit trail every morning while you are learning it. Ten minutes is enough.
- Get your own numbers in order. Real stock across every channel in an inventory system, and real landed costs in your accounts, so any pricing rule you give an AI is built on your figures rather than guesses.
If you are already using Claude elsewhere in the business, our guide to Claude's small business workflows covers the connections that are worth setting up alongside this one. And if you want help deciding what an AI should and should not be allowed to do in your own operations, talk to us. Deciding that line for our own always-on systems is most of what we do.
