Sometime in the next few weeks, probably during a keynote in Seattle, you are going to hear that Amazon's own AI can now run your seller account. Amazon Accelerate runs September 22-24, and everything Amazon has shipped this year points at one message: Seller Assistant is no longer a help-page chatbot. It monitors your inventory, rewrites your listings, scans your compliance, generates your ad creative, and takes approved actions on its own. It is free, it is already in your Seller Central, and it raises the question every seller will be asking by October: if Amazon gives me an AI employee for free, do I still need my own?
Short answer: you need both, because they work for different bosses. The long answer is worth ten minutes, because the two kinds of AI are good at genuinely different jobs, and the sellers who get 2027 right will be the ones who gave each assistant the work it is actually built for.
What Seller Assistant actually is now
The facts, dated, because this space moves fast. Amazon announced the agentic Seller Assistant at Accelerate in September 2025 and rolled it out to all US third-party sellers by December 2025. It runs on Amazon Bedrock using a combination of Amazon Nova and Anthropic Claude models. It is free, with no premium tier - every Seller Central account has it. As of mid-2026 its core abilities are:
- Inventory management - monitors FBA levels, flags slow movers before long-term storage fees land, and recommends specific SKUs to mark down or remove.
- Enhance My Listing (EML) - proposes new titles, bullets and descriptions, including suggestions aimed at Alexa for Shopping.
- Compliance monitoring - scans listings against policy and safety requirements and alerts you to missing documentation before a suppression does.
- Creative Studio - generates ad creative from conversational prompts.
- Demand forecasting - reads your sales patterns and prepares shipment recommendations.
- Business intelligence - natural-language questions about your account, with Canvas (US and UK since March 2026) building interactive dashboards from a prompt.
Actions run through a three-tier permission model: suggest-only, auto-approve for low-risk categories, and manual-only for pricing and live listing changes. Amazon's own line is that the agentic system takes only approved actions.
Where it is genuinely good - use it
Let us be plain about this before the caveats: Seller Assistant is a real gift, and turning it on costs you nothing. Compliance monitoring alone justifies it - a free scanner that flags a missing certification before the listing gets suppressed is pure upside. The inventory alerts catch slow movers you stopped watching. Creative Studio removes a real cost for sellers who were paying per ad image. And it reads Amazon-internal signals - policy state, safety requirements, fulfillment machinery - that no third-party tool sees directly. If you do nothing else after reading this, open Seller Central and switch it on.
The three walls it cannot see past
None of what follows is a bug Amazon will fix, because none of it is a bug. It is what Seller Assistant is for.
Wall one: it works for the marketplace, not for your P&L. Seller Assistant does not know your product costs, and it does not ask. It can tell you sales are up; it cannot tell you whether you made money, because true profit lives in numbers Amazon never holds - what you paid for the goods, your inbound freight, your overhead. Every recommendation it makes is optimized inside Amazon's frame: sell more, stock correctly, comply. Those goals usually align with yours. When they do not - when the right move for your margin is to sell less of something, raise a price, or let a low-margin listing starve - a marketplace-owned assistant is structurally the wrong adviser to ask.
Wall two: it lives where your data ends. One account, one marketplace view, no context from anywhere else. Canvas dashboards cover the US and UK; cross-marketplace forecasting is not there. And the well-documented weak spot: Enhance My Listing writes copy without your keyword performance context. September 2026 coverage of the tool is blunt about it - accepted Enhance My Listing suggestions can drop phrases that were carrying organic rank, with ranking losses visible within days. The fix every guide recommends is the same: cross-reference its output against your own search-term data before accepting anything. Which is another way of saying it needs an assistant of its own.
Wall three: it lives in Seller Central. Your thinking probably does not happen there. It happens in the AI you already talk to all day, where you draft plans, compare options and ask follow-up questions for as long as it takes. Seller Assistant answers inside Amazon's walls, one account at a time, in Amazon's interface.
The other kind of assistant
The complement is an AI that works for you: the Claude or ChatGPT subscription you already have, connected to your own data. This is what we build, so weigh the source - but the architecture matters more than the vendor. Through an MCP connection (the open standard both Claude and ChatGPT speak), AMZ Vault connects your Amazon account to your own AI: your full P&L with your real product costs, every fee and settlement, PPC down to the search term, inventory, forecasts. Your AI can read all of it, analyze anything, and stage changes - bids, budgets, negatives, listings - that wait in an approval queue. Nothing reaches Amazon until you click Approve.
That flips all three walls. It optimizes for your profit because it knows your costs. It sees across your whole business rather than one console. And it works wherever you think - including on a schedule, unattended, from either Claude or ChatGPT.
Who should own which job
The honest division of labor, as of September 2026:
| Job | Give it to | Why |
|---|---|---|
| Compliance and policy alerts | Seller Assistant | Reads Amazon-internal signals no outside tool sees; free; low-risk auto-approval fits it |
| Ad creative generation | Both | Creative Studio is free and native to the ad system; generating your own is more capable and builds on your own brand assets and data. Use whichever you are most comfortable with |
| Slow-mover and storage-fee flags | Both | Take Amazon's alert, then ask your own AI whether the markdown clears your cost or just Amazon's shelf |
| Listing copy changes | Your own AI | Yours writes from your keyword and search-term data - the exact context EML lacks. EML drafts are a fine starting point, but check every dropped phrase before accepting |
| Profit questions | Your own AI | Requires your costs; Amazon does not hold them |
| PPC bid, budget and negation decisions | Your own AI | Decisions should clear your break-even economics, not engagement metrics, and every change should keep an approval trail |
| Reorder and cash decisions | Your own AI | Amazon's forecast optimizes stock availability; yours should also weigh cash, margin and supplier terms |
Capabilities and availability as of September 2026, before Accelerate; Amazon may expand the left column at the keynote, and we will update this post when it does.
The takeaway
Do not frame this as Amazon's AI versus yours. Frame it as two employees: one supplied free by your landlord, excellent at the landlord's paperwork and genuinely useful around the building - and one on your payroll, the only one allowed to see your books. Turn Seller Assistant on this week. Then give the decisions that move your money to an AI that knows what your money looks like - and make sure anything either of them wants to change waits for your approval first.
Related: Amazon's AI Agent Policy, explained · every Amazon MCP server compared · try the whole platform with no signup