Amazon's Agent Policy: what AI tools and agents may do on your seller account
Amazon's Agent Policy sets rules for any software that acts autonomously or semi-autonomously on an Amazon service, AI or not. Section 19 of the Business Solutions Agreement made it binding on March 4, 2026. Agents must name themselves in every request's user agent string, never pose as a person or solve CAPTCHAs, answer truthfully, respect Amazon's limits and stop when asked. AI tools are not banned.
- Amazon announced the Agent Policy in the Seller Forums on February 17, 2026; it took effect with the updated Business Solutions Agreement on March 4, 2026.
- Section 19 requires every Agent to identify itself as an automated system, follow the Agent Policy at all times and stop accessing Amazon's services if Amazon explicitly asks.
- The Agent Policy requires 'Agent/[agent name]' in the user agent string of every HTTP or HTTPS request, and bans mimicking human keystrokes or solving CAPTCHAs.
- The agreement defines an Agent as any software or service that takes autonomous or semi-autonomous action for a person or entity, so the rules are not limited to AI.
- Section 4.2 bars using Amazon's materials or services to develop or improve large language, multimodal or machine learning models, or letting a third party do so.
- Where Amazon's confidential information reaches the SP-API developers a seller authorizes, the seller stays liable for them as if their acts were the seller's own.
This post works from Amazon's own text, not from summaries of it: the Business Solutions Agreement (BSA) that every seller accepts, the one-page Agent Policy in Seller Central Help, and the Acceptable Use Policy that SP-API developers must follow, all checked September 25, 2026. We have built on Amazon's seller APIs since 2017, so after the rules we explain what each one means for the software connected to your account, and what to ask before you connect more.
What is Amazon's Agent Policy?
The Agent Policy is a short list of rules for any Agent that accesses, uses or interacts with an Amazon service. Amazon publishes it in Seller Central Help under Program Policies, and Section 19 of the BSA makes it part of your contract: all Agents must "clearly identify themselves as automated systems" and follow the policy "at all times." The BSA defines an Agent as "any software or service that takes autonomous or semi-autonomous action on behalf of, or at the instruction of, any person or entity." The Agent Policy repeats that definition in its first sentence.
That definition is wider than "AI". Read literally, it covers a large language model that edits listings, and also a rule-based repricer, an ad-bid script or a browser extension that clicks through Seller Central for you. An AI assistant connected to your account through an SP-API MCP server acts at a person's instruction, so on the same reading it counts too. Agents that shop on behalf of buyers are a separate topic, covered under agentic commerce.
Take a tool we built ourselves. Repriser is rule-based automation we built for a client: when a competing offer changes, it reprices within the rules the client set, without a person approving each change. We do not call it an AI agent, because no step uses a model. Amazon's definition does not ask that question. Software that acts on its own on your account appears to be an Agent in Amazon's sense, whether it runs on rules or on a model.
When did it take effect, and who does it apply to?
Amazon announced the change in a Seller Forums post, Business Solutions Agreement updates effective March 4, 2026, published on February 17, 2026 (checked September 25, 2026). The same update added limits on using Amazon's materials for AI development, a separate agreement for the Mexico store, a new Section 20 on dispute resolution, and definitions of "Agent" and "Our Materials." There was no acknowledgment step: the post says continued use of Selling Services after the effective date "constitutes acceptance of these changes," and Section 15.3 of the BSA says the same.
The Agent Policy applies whenever you "use, allow, enable, or cause the deployment of" an Agent, and Section 19 whenever you "deploy, enable, or authorize" one. "You" is the business that accepted the BSA, so the wording reaches tools an agency or a freelancer runs on your account, not only the ones you install yourself. The BSA's Services include Selling on Amazon, FBA, Amazon Advertising and the Selling Partner APIs, so we read the rule as reaching all of them.
Developers are bound from their side too. Section 1.3 of the Solution Provider Portal Agreement, the agreement developers accept in Amazon's developer portal, repeats the same Agent rules and points to the Agent Policy for the technical standards (checked September 25, 2026). The Acceptable Use Policy adds that developers must not "facilitate or promote violation" of a seller's agreement with Amazon, and it names the BSA.
| Who | How the rule reaches them | What to do now |
|---|---|---|
| Brand owner (the seller) | Section 19 of the BSA and the Agent Policy apply when your business uses, allows, enables or authorizes an Agent | List every app, user and extension with access, then run the checklist below on each |
| Agency or freelancer working in your account | Their tools act on your account with your permission, so your BSA covers them; an agency with its own SP-API app is also bound by Section 1.3 of the developers' agreement | Put the checklist answers in the contract |
| App developer or SaaS vendor | Section 1.3 of the developers' agreement, the Agent Policy, and the Acceptable Use Policy's bar on helping a seller break its agreement | Identify the agent in every request and document its data access, limits and stop procedure |
What must an AI agent or automated tool do?
Section 19.1 of the BSA sets three duties: identify as an automated system, follow the Agent Policy "at all times," and do not access Amazon's services once Amazon "has explicitly requested cessation." Section 19.2 adds that Amazon may limit or restrict agent access "through technical or other measures," at its sole discretion. The Agent Policy then lists four rules. An Agent must:
- Say it is an Agent, and give its name, in every HTTP or HTTPS request, by adding
Agent/[agent name]to the user agent string. Amazon's own example isAgent/AmazonAgent. - Never hide that it is an Agent, for example by mimicking the speed or pattern of a person's keystrokes and page navigation, or by completing or circumventing CAPTCHAs.
- Answer truthfully when asked whether the interaction comes from a person or a computer.
- Never get around any measure Amazon uses to block, limit, modify or control how Agents use its services.
Here is what the first rule means in practice. For software built on the SP-API, Amazon's Connect to the SP-API guide already requires a user-agent header on every request, up to 500 characters, with at least the app name, version and programming language (checked September 25, 2026). The Agent Policy asks for Agent/[agent name] in that same user agent string. Amazon's pages do not show a combined example, so ask your developer for a real request header from their logs, not a promise.
What is not allowed?
The Agent Policy and Section 19 already rule out the tactics that let software pass as a person: human-like typing and clicking, CAPTCHA solving, false answers about being human, working around blocks or rate limits, and carrying on after Amazon says stop. The Acceptable Use Policy separately bars developers from creating multiple accounts or apps to get around throttling, and the BSA's Selling Partner API terms bar you from letting anyone use the APIs in a way intended to avoid "exceeding usage limits or quotas." Section 4 of the BSA then adds three limits that matter for any AI tool:
- Training AI on Amazon's data. Section 4.2 bars using Amazon's materials or services "to directly or indirectly develop or improve" large language, multimodal or machine learning models. It says "you will not, and will not allow any third party to," so it covers your vendors. The BSA defines "Our Materials" to include the "data, materials, and other items or information" Amazon makes available to you, which, read plainly, includes your reports and API data.
- Scraping. The same section bars "data mining, robots, or similar data gathering and extraction tools" for collecting Amazon's materials.
- Reverse engineering. Section 4 also bars reverse engineering Amazon's materials, including to derive "a model, model parameters, or model weights."
On our reading, the training clause targets building or improving models, not asking a model a question about your own sales. The gray area is a vendor, or the AI provider behind it, that keeps what it is sent and uses it to train. That is why the checklist below asks which model provider a tool uses and what its data terms say, and why the answer belongs in the contract rather than in a sales call.
What are the risks for the seller?
The seller carries most of the risk, because the agreement is between Amazon and your business, whoever wrote the code. The BSA's Selling Partner API terms say that where Amazon's confidential information is disclosed to developers you authorize, you remain liable for them "as if such acts or omissions were your own." Its password clause makes you "solely responsible for any use of or action taken under your password." Section 19 names no penalty of its own, so the general terms apply.
| Risk | What Amazon's text says | How to reduce it |
|---|---|---|
| Amazon restricts an agent | Section 19.2: Amazon may limit or restrict agent access at its sole discretion | Keep a manual fallback for every job an agent does, such as repricing, bids and reorders |
| Your account is suspended | Section 3: suspension or termination for a material breach not fixed within 7 days of a cure notice, or at once if your use "has harmed, or our controls identify that it might harm" other sellers, customers or Amazon's legitimate interests | Connect only tools that pass the checklist below, and remove the ones you no longer use |
| A vendor's mistake becomes yours | You stay liable for authorized developers; you answer for actions taken under your password | Never share your login; give each person their own user with limited permissions |
| The tool stops working mid-season | The Acceptable Use Policy says breaking it may lead to suspension or termination of a developer's API access | Ask vendors how they track Amazon's policy changes, and what happens to your data if access ends |
Questions to ask any AI tool or agency before you connect it
Each row maps one line of Amazon's text, quoted and linked, to a question for the vendor and the kind of answer that should satisfy you. The same questions work for a SaaS app, an agency's in-house tool or a developer you hire. All sources checked September 25, 2026.
| What Amazon's text says | Ask the tool or agency | A good answer |
|---|---|---|
Put Agent/[agent name] in every request's user agent string (Agent Policy, rule i) | What agent name does your software send, and can we see one logged request? | One fixed agent name on every request, shown in a real request log |
| No "mimicking the speed or pattern of human keystrokes," and no solving CAPTCHAs (Agent Policy, rule ii) | Does your tool ever drive Seller Central through a browser or an extension, and what does it do at a CAPTCHA? | It works through Amazon's APIs; at any verification prompt it stops and alerts a person |
| "Respond truthfully" when asked if it is a person or a computer (Agent Policy, rule iii) | If Amazon asks whether an action came from a person or software, what does your system answer? | Software, every time |
| Never get around measures that "block, limit, modify, or control" agents (Agent Policy, rule iv) | What happens when Amazon throttles or blocks a request? | It slows down within Amazon's rate limits and tells you; no extra apps, accounts or proxies |
| No access once Amazon "has explicitly requested cessation" (BSA, Section 19.1) | If Amazon asks your agent to stop, how is it stopped on our account, and who tells us? | A pause for your account alone, a named person who triggers it, and your own right to revoke the app |
| Do not use Amazon's materials to "develop or improve" AI or machine learning models (BSA, Section 4.2) | Does any data from our account train or improve your models, or your AI provider's? | No, in the contract, with the model provider and its data terms named |
| Developers may not "request or share Amazon Portal usernames or passwords" (Acceptable Use Policy, 3.3) | Do you need our Seller Central login or our verification codes? | No. You authorize the app in Seller Central, and people get their own user through secondary user permissions (3.7) |
| Do not request data that is "not necessary" for the app to work (Acceptable Use Policy, 3.8) | Which SP-API roles does your app request, and why each one? | A short list tied to features, with buyer personal data only for seller-fulfilled shipping or legal needs, the only uses the Acceptable Use Policy allows (4.1) |
| Be explicit about "the use of models such as artificial intelligence" (Acceptable Use Policy, 2.4) | Which steps follow fixed rules, which use an AI model, and how fresh is the data? | A written split of rule steps and model steps, with data refresh times |
| You stay liable for developers "as if such acts or omissions were your own" (BSA, Selling Partner API terms) | Which actions does the agent take without our approval, and where is the log of every change? | A written list of actions within your rules, approval for money and brand decisions, and a change log you can export |
You can check one thing yourself today. In Seller Central, open Apps and Services, then Manage Your Apps, to see the apps connected to your account. Amazon's Revoke Authorizations guide says a seller removes one by choosing Disable authorization, after which it stays on the page as disabled. Its Seller Authorizations FAQ adds that Amazon's routine clean-up removes authorizations that have gone more than 365 days with no renewal and no API call, so an app nobody remembers can stay connected for a year or more (both checked September 25, 2026).
How do we design agents to follow the policy?
Ecomsellertool is a tech agency that grows brands through technology. We build and deploy AI agents on a brand's own Amazon, Walmart and Shopify accounts, on top of Ecomsellertool Growth OS, and we have built on Amazon's seller APIs since 2017. Our agents are designed to follow Amazon's Agent Policy, and the design starts from principles we already publish on our AI agents page.
Each agent works through Amazon's APIs on accounts you authorize. Routine actions can run within the rules you approve. Budget changes, purchase orders and unusual cases can come back to your team with the evidence and a proposed next step. Each agent page lists which steps follow fixed rules and which use an AI model. The agents run on your accounts, not ours. You keep your accounts, your data and the custom code we build; the Growth OS base is licensed to you.
This post does not list code-level controls, because they differ from one agent to the next. Each custom agent is scoped in an architecture doc with a fixed price and go-live date before work starts. That is where the checklist answers belong, in writing: the agent name it sends, what it does on its own, what comes back to your team and how it is stopped. Ask us the same questions you would ask anyone else.
You may not need a custom agent at all. If all your work sits inside Amazon, Amazon's own Seller Assistant may be enough. Amazon says it works in Seller Central, Amazon Quick and Claude, and that sellers review and approve its actions before they are carried out; the plugin for Quick and Claude is in beta for sellers in Amazon's US stores (About Amazon, September 23, 2026, checked September 25, 2026). A custom agent makes sense when the work has to act on Walmart, Shopify, your 3PL or your ERP as well as Amazon, or run on your own accounts and code. Our Seller Assistant plugin explainer covers what Amazon's tool reaches.
If you are building a tool yourself with an AI coding assistant, our guide to building Amazon seller tools with AI covers the SP-API rules those assistants tend to miss. Before any tool or agency connects, give it access the safe way. And for which tasks are worth automating at all, see how to automate Amazon operations.
Where should you start?
Run this review once now, and again whenever you add a tool, an agency or a freelancer.
- In Seller Central, open Apps and Services, then Manage Your Apps, and list every app connected to the account.
- List every person and service provider with access in your user permissions. Our guide to giving an agency access to Seller Central safely covers which permission each one should get.
- Ask your team which browser extensions or scripts they use inside Seller Central.
- Put each app, agency and extension through the checklist above, and disable anything nobody can explain.
- Keep a manual fallback for repricing, bids and reorders, in case Amazon restricts an agent.
If you want to see where software could take routine work off your team within these rules, start with the free 24-hour diagnostic: connect Amazon with Login with Amazon, no password shared, and get a report within 24 hours of connecting, on business days. We only read data; we never change listings, prices, stock or ads. If you are choosing an agency, our agency RFP questions put these checks into a wider template.
Frequently asked questions
Are AI tools still allowed on Seller Central?
Yes. Neither the Agent Policy nor Section 19 of the Business Solutions Agreement bans AI tools, repricers or other automation. They set conditions: the software must identify itself as an Agent in every request, must not pose as a person, must respect the limits Amazon sets and must stop if Amazon asks. Amazon also keeps the right to limit or restrict agent access at its sole discretion, so any tool's access can change.
Does a rule-based repricer or bid tool count as an Agent?
Read literally, it appears to. The agreement defines an Agent as any software or service that takes autonomous or semi-autonomous action on behalf of, or at the instruction of, any person or entity. That wording does not mention AI, so repricers, bid rules and scripts that change your account without a person making each change seem to be covered. This is our reading of Amazon's text, not legal advice.
Does the Agent Policy cover SP-API apps, or only bots that use the website?
Amazon's text does not limit it to the website. The Agent Policy covers any Agent that accesses, uses or interacts with any Amazon Service, and the Business Solutions Agreement lists the Selling Partner APIs among its Services. Developers also accept the same Agent rules in Section 1.3 of the Solution Provider Portal Agreement. On our reading, an app that acts on your account through the SP-API should identify itself as an Agent too.
Did I have to accept the new agreement?
There was no separate step. Amazon's February 17, 2026 announcement says continued use of Selling Services after the March 4, 2026 effective date constitutes acceptance of the changes, and Section 15.3 of the agreement says the same about any change. If a change is unacceptable to you, the agreement says to stop using the services and end it.
Can an AI tool ask for my Seller Central password?
It should not. Amazon's Acceptable Use Policy tells SP-API developers not to request or share Amazon Portal usernames or passwords, and to ask for access through secondary user permissions when a person needs to work in Seller Central. The Business Solutions Agreement also makes you solely responsible for any action taken under your password.
How do I disconnect an AI tool from my Amazon account?
In Seller Central, go to Apps and Services, then Manage Your Apps, find the app and choose Disable authorization. Amazon's SP-API documentation says the app then stays listed on that page as disabled. Also remove any user or service provider access you gave the vendor, and change any password you shared.
Does the Agent Policy apply outside the US?
This post quotes the US Seller Central versions of the agreement and the policy. Amazon's announcement covered the US and Canada agreement and a new, separate agreement for the Mexico store. If you sell in other Amazon stores, read the agreement and policies in each store's Seller Central before you connect a tool there.
How we research, fact-check and compare: our editorial standards. Spot an error? Email hello@ecomsellertool.com and we will correct it.
