Amazon has moved Seller Assistant across the line that separates advice from action. It now remembers a seller’s business between sessions, and it can run workflows that watch market conditions and respond to them without anyone logged in. Amazon has also opened a plugin that lets outside AI tools read Seller Central data, launching first in Amazon Quick and in Anthropic’s Claude.
Read alongside Amazon’s decision to shut Meta’s shopping agent out of its store this same month, a pattern appears. Amazon is open when it is the supplier of data and closed when it is the shop.
What changed
The memory is the first change. Seller Assistant “retains a persistent memory of pricing patterns, inventory cycles, and growth goals”, which is what turns a question-and-answer tool into something that can make a recommendation with context behind it.
The second change is the one that matters operationally. The new workflows “monitor market conditions and respond dynamically around the clock, even when the seller isn’t logged into Seller Central.” Amazon describes seller-defined guardrails and audit trails around them. Alongside this, inbound planning and aged inventory agents are being added inside the same assistant.
An assistant that suggests a price change is a productivity tool. Software that changes the price at three in the morning because a competitor moved is a different category of thing, and it should be governed differently.
From the publisher
Everything we know about running Amazon, in one course.
The commercial strategy, taught by operators who run our own brands.
Three different 90 percents
The coverage of this announcement contains at least three separate figures of 90 percent, and they are not interchangeable. Quoting any of them without its label produces a claim nobody made.
The first is adoption of the assistant: Amazon says it has rolled out to over 90 percent of its selling partners worldwide in their native language. The second is behaviour: sellers accept its recommendations more than 90 percent of the time, reported as a 90 percent recommendation acceptance rate against hundreds of thousands of active users. The third is about other software entirely: 90 percent of selling partners already use third-party AI tools.
Availability, acceptance and third-party usage are three different measurements that happen to round to the same number. The acceptance rate is the one most likely to be repeated and the one that deserves the most suspicion, because a recommendation a seller can accept with one click and ignore with none is not a hard test, and Amazon has published no base rate for comparison.
Tool comparison · FastMoss vs Kalodata
An assistant that reprices against conditions inside one platform cannot see what is happening on another. Sellers running more than one channel still need a read on demand and pricing that the platform itself does not provide. FastMoss and Kalodata both cover TikTok Shop product and pricing data independently.
Comparison coming soon
FTC disclosure: we may earn a commission if you buy through links on this page. It costs you nothing and does not change what we recommend.
Where each part is available
These do not share a geography and merging them produces a wrong claim. The Seller Assistant upgrades are described as rolled out to over 90 percent of selling partners worldwide, which means a European seller can go and look at them.
The plugin is different. Amazon describes it as in beta for sellers in its United States stores, with international expansion to follow. A European seller cannot connect Claude to Seller Central today, and anyone describing that as a live European capability has merged two announcements.
The Amazon Quick offer is the broadest of the three: every primary account holder globally can sign up for a free 12-month Amazon Quick Plus subscription, available through the end of 2026.
The inconsistency worth naming
Within the same month, Amazon has done two things that sit oddly together. It asked Meta to remove Amazon from a shopping agent and enforced that on its own site, on the principle that agents acting for customers should respect a service provider’s decision about participation. And it built a plugin that exposes its own sellers’ operational data to outside AI tools.
Both positions are defensible on their own. Together they describe a company that welcomes being read and refuses being transacted with. The line is not about artificial intelligence at all. It is about who owns the purchase, and Amazon is consistent on that even where it looks inconsistent on everything else.
For a seller, the practical read is that Amazon will keep opening data interfaces and keep closing buying ones, and that any strategy built on agents transacting on Amazon is building against the grain.
What unattended action actually requires
The guardrails are the product. A workflow that reprices without supervision needs a floor price that reflects landed cost rather than last year’s cost, a ceiling that stops it chasing a competitor out of a category, and a review cadence that catches drift.
Most sellers do not have a current landed cost per unit. That is the honest blocker here, and it is not a software problem. Automation applied to a margin figure nobody has checked since the last freight change does not save time, it spends money faster.
European sellers have a particular reason to check that figure this autumn rather than next spring. Landed cost is moving twice in quick succession: monthly storage in the European fulfilment centres steps up for the October to December quarter, and an EU-wide handling fee on imported parcels arrives in November with its amount still unset. A floor price calculated in the summer is already describing a cost base that no longer exists.
The aged inventory agent is the one to watch for the same reason. Its job is to flag stock that has sat long enough to attract surcharges, and those surcharges are at their most punitive in exactly the quarter the tool is arriving. That is a useful coincidence rather than a coordinated one, but it makes this the most valuable time of year for that particular alert to start firing.
Tool comparison · FastMoss vs Kalodata
Setting a sane floor and ceiling for automated pricing means knowing what the category actually sustains, not what one competitor did yesterday. Category-level pricing history is the input most repricing setups are missing. FastMoss and Kalodata both provide it for TikTok Shop.
Comparison coming soon
FTC disclosure: we may earn a commission if you buy through links on this page. It costs you nothing and does not change what we recommend.
For the business operator
Go and look at what Seller Assistant is already recommending before switching anything to run unattended. The recommendations are visible now in European accounts, and reading a fortnight of them tells you whether the suggestions would have been good ones.
Before enabling any always-on workflow, write down the floor price for your top twenty products using a landed cost calculated this quarter. If that number does not exist, the automation is not the next task.
Treat the audit trail as the thing you are buying. An automated action you cannot reconstruct three weeks later is worse than no automation, because it removes the ability to learn from the outcome.
From the consumer’s side
Automated repricing is already why an item a shopper saw yesterday costs something different today, and more sellers acting continuously will make prices move more often rather than less.
The consumer consequence is a loss of trust in the price rather than a loss of the sale. Shoppers who notice a price rising and falling within a day stop believing any of the numbers, start waiting, and become considerably harder to convert with a discount, because they have learned that the discount is weather rather than an offer. A seller whose prices move less than the category can say so, and that is worth more in a market where everything else is twitching.
The three-line version
If you run pricing or marketplace operations, read a fortnight of Seller Assistant recommendations before delegating anything to it, and do not enable an unattended workflow without a floor price built on this quarter’s landed cost. The line to take upward is that the capability is real and available in Europe, and the blocker is our own cost data rather than the tool.
If you have just taken on e-commerce operations, the gap this exposes is governance. Nobody in most organisations owns the question of what software is permitted to change without a human, and that question is now live rather than theoretical.
If you sign the budget, this one is free, which is the point. Seller Assistant costs nothing and the Amazon Quick subscription costs nothing through the end of 2026. Refuse any proposal to buy a third-party repricing tool this quarter until somebody has tested what the free one already does.
Sources
ChannelX — https://channelx.world/2026/09/amazon-seller-assistant-plugin-for-amazon-quick-and-anthropics-claude/
Supply Chain Dive — https://www.supplychaindive.com/news/amazon-ai-supply-chain-agents-among-seller-upgrades/831164/
Retail Dive, on Amazon and Meta’s shopping agent — https://www.retaildive.com/news/amazon-meta-muse-ai-agent-shopping-experience/831018/