AI Agent Rules for Sellers: A Do Alone, Ask First, Never Template

AI agent rules for sellers come down to three lists: what the agent may do alone, what it must ask you about first, and what it must never touch. An always-on agent that can see your listings, prices and buyer messages will act on whatever you leave unwritten, and the expensive mistakes (a price typed one digit short, a refund sent on the wrong order, a polite message asking a buyer to update a one-star review) take seconds to make and weeks to clean up. Write the three lists before the agent gets access, not after the first bad night.

This post is the hands-on part. It gives you a copy-paste template across listings, pricing, ads, refunds, customer messages and reviews, and inventory. It marks which rows come from a written marketplace rule and which are only our suggested default, then maps the result onto OpenAI's Custom Rules. If you are still deciding whether to hand anything to an agent, start with what the product can and cannot do in the guide to OpenAI Dots for sellers.

Also ReadCan an AI Agent Run Seller Central? What Amazon and Shopify Allow

›

Why does an always-on agent on my store need written rules first?

An always-on agent needs written rules first because it keeps working while you sleep, and what it does under your account is treated as something you did. Shopify's Terms of Service say: "You will be responsible for any breach of the Terms of Service by your affiliates, agents or subcontractors." On Amazon, an Amazon staff reply in the Seller Forums about review rules says "This statement includes anyone associated with your account", and an agent you run is associated with your account.

The second reason is speed. A human assistant who is unsure usually stops and messages you. An agent that meets a vague instruction may pick the most likely reading and carry on, then repeat that reading across forty listings before you look. Written rules turn "use your judgment" into a short list it checks before each action, and they give you something concrete to audit against at the end of the week.

Rules sit on top of the platform's own limits, not in place of them. OpenAI's controls page says "Built-in safeguards, existing app permissions, and automatic approval checks apply from the start." It also says rules do not grant access to an app or computer, do not override built-in safety requirements and do not remove required confirmations, such as approval to use a saved login. A rule can make an agent more careful than the defaults. It cannot make a forbidden marketplace action acceptable.

What should AI agent rules for sellers cover: do alone, ask first or never?

AI agent rules for sellers should sort every task into one of three buckets: do alone, ask first or never. Do alone covers work that only reads data or produces a draft nobody sees until you approve it. Ask first covers anything that changes what a buyer sees or pays, or that spends your money. Never covers the short list of actions a marketplace rule forbids, plus anything that needs your password.

Also ReadThe Amazon Listing Audit: Five Passes Instead of Random Tweaking

›

• Do alone: reversible, invisible to buyers, no money moves. Audits, reports, keyword lists, drafts.
• Ask first: visible to buyers or moves money. Live listing edits, prices, bids, refunds, buyer replies, purchase orders.
• Never: a marketplace rule says no, or the task needs a password. The agent hands it back to you.

Start with ask first as the default for everything, including the tasks you expect to move to do alone later. A week of approving actions one by one is tedious, but it shows you exactly where the agent reads your instructions differently from how you meant them. After seven days, move a task to do alone only if every instance that week was correct with no edits from you. A task that needed even one correction stays in ask first for another week.

What is the copy-paste rules template for my store?

The copy-paste template below turns these AI agent rules for sellers into one table, listing each task once, with the bucket it belongs in. Never rows marked "Marketplace rule" come from an Amazon or Shopify source quoted in the next section. Rows marked "Our suggested default" are our recommendation for a small seller starting out. We did not find an Amazon page that sets rules for agent bid changes, refund approvals or reorders, so treat those rows as a starting point you adjust to your own margins and volume.

Paste the table into the agent's instructions as plain text, one line per row, with the never rows at the top so they are read first. Keep the wording literal: "Never offer a buyer anything in exchange for a review" is easier for an agent to follow than "be careful with reviews". If you sell on more than one marketplace, keep one copy per store, because a reply rule that suits Amazon buyer messaging may not match how you handle Shopify customer email.

Which marketplace rules turn some tasks into hard never rules?

Five tasks are hard never rules because a marketplace has said no in writing: offering anything for a review, asking a buyer to change a review, sending buyers marketing messages, sharing passwords and sending spam. On reviews, an Amazon staff reply in the Seller Forums says "Amazon has a zero-tolerance policy towards customer reviews manipulation." It lists "Offering incentives (discounts, free products, refunds, or other compensation) in exchange for a review." and "Requesting changes to existing reviews, whether through Buyer-Seller Messaging, product inserts, or external channels." Read both lines in the Amazon Seller Forums review thread.

Buyer messages have their own limit. In a separate Amazon staff reply in the Seller Forums, Amazon says "In general, you can contact buyers only to complete orders or to respond to customer service questions." The same reply says "You cannot contact buyers for marketing or promotional purposes" and that "Amazon has the authority to block any message at its discretion." An agent that writes friendly follow-ups about your new product line is sending exactly the kind of message that reply rules out.

Also ReadOpenAI Dots for Amazon and Shopify Sellers: What You Can Safely Hand Off

›

Passwords are the third hard line. Amazon's developer Acceptable Use Policy says "Never share access keys or passwords." It is written for developers rather than sellers, but it is a sound rule of thumb for an agent as well. An Amazon staff reply in the Seller Forums also says "Don't add external third-party service providers or developers as users.", and Shopify's Terms say "You are responsible for keeping your password secure." Whether an agent may sign in to Amazon at all is a separate question, covered in the guide to AI agents in Seller Central.

The last hard line is spam. Shopify's Acceptable Use Policy rules out "using Shopify as a platform to send spam, or for other malicious practices", so an agent should never send bulk unsolicited email from your store. One more Amazon note applies if an agent ever deals with Amazon systems: an Amazon update on agents posted in the Seller Forums says they should "Clearly identify themselves as automated systems". Treat that as a reason to keep the agent honest about what it is, never as permission to act.

What can an agent do alone on listings, and what must it ask about prices, ads and inventory?

An agent can do listing work alone when the output is a draft or a report: audits for missing attributes, empty image slots and title length, keyword research, image alt text and rewritten bullets saved as drafts. None of that changes what a shopper sees until you publish it, and a bad draft costs you a minute of reading rather than a confused buyer. Keyword lists are still worth a second look against an independent source, such as an Amazon keyword generator.

Publishing is where the bucket switches to ask first. Before the agent pushes any edit live, have it show the current text, the proposed text and the reason, side by side, in one message you can approve or refuse. If Amazon has already rewritten a title on its own, tell the agent to report it to you rather than change it back, so the two of them do not keep overwriting each other.

Prices, coupons and shipping charges belong in ask first, with no exceptions in the first month. An Amazon staff reply in the Seller Forums notes that "Amazon regularly monitors the prices of items on our stores, including shipping costs", so a careless repricing run is visible to Amazon as well as to your buyers. Our suggested default is to let the agent watch prices and propose changes inside a floor and a ceiling you set per SKU, and to approve each change yourself.

Ads follow the same pattern. Let the agent pull search term reports and flag keywords with spend and no sales, and make every bid, budget or campaign change an ask first item. Once it has earned some trust, our suggested default is a one-way cap: the agent may lower a bid on its own, but it never raises a daily budget. We did not find agent-specific ad rules from Amazon, so this is a spending safeguard, not a policy.

Refunds and inventory are money leaving your account, so both stay in ask first. The agent can draft a refund recommendation with the order ID, the buyer's message and its suggested action, and it can watch stock and propose a reorder date and quantity. You approve or change the number. Keep refund decisions with a person for as long as you can, because a refund is one of the incentives Amazon lists when it describes review manipulation.

How do I turn the template into OpenAI Custom Rules?

To turn the template into OpenAI Custom Rules, map each bucket onto one of the four behaviors OpenAI offers, keeping in mind that never is not one of them. OpenAI's own advice is to start simpler: "Start with clear instructions in your conversation; you don't need to create a rule for every approval." Paste the template as instructions first, then add Custom Rules for the rows where a slip would cost real money.

The hand-off behavior is the closest thing to a never rule: the agent does not do the task, it gives the task back to you. That keeps the decision with a person, which is what you want for anything a marketplace forbids. Rules have limits in the other direction too, because they cannot grant an agent access it does not already have or switch off the confirmations OpenAI requires, as the OpenAI Dots controls page explains.

Expect confirmations even on rows you set to do alone. OpenAI says "Reusing a saved login to sign in requires your confirmation." Its controls page also says an automatic review "determines whether the action can proceed, needs your approval, or includes a step you must do yourself", and gives the example "For example, you must change a password yourself." If you run the agent in Slack, its channels page says "By default, your dot responds to you", so check that setting before you add teammates to the channel.

How do I audit what the agent did and stop it fast?

Audit the agent by checking results, not run logs, because OpenAI says "A completed run doesn't by itself confirm that the requested result was achieved or delivered." For each action in the week, open the listing, the order or the ad report and confirm the change is really there and really correct. If the agent repeats work on a schedule, OpenAI's advice is "Ask your dot to confirm what it saved." so you know exactly what runs next time.

• Day 1: paste the template as instructions, with every task set to ask first.
• Days 1 to 7: approve or refuse each request, and write one line on why you refused.
• Every evening: check three changed items against the live store, not against the agent's summary.
• Day 7: move a task to do alone only if it had zero corrections all week.
• Every week after that: rerun the same listing check and compare it with the week before.

For the listing side, an independent check catches what a tired eye misses at the end of a long day: a title over your length limit, a missing bullet, an image slot left empty or a backend field the agent never filled. Run the agent's drafts and your live pages through the same checklist every week, using a tool such as an Amazon listing quality checker.

Stopping an agent is less complete than it looks. OpenAI's controls page says "Pause stops your dot's current main task. It doesn't stop every delegated task or cancel future scheduled runs." It adds "Stopping work doesn't undo completed actions." So your kill switch has three steps: pause, cancel the scheduled runs, then reverse any completed changes by hand, starting with prices and live listings because buyers see those first.

Keep a short change log while you do this: date, task, what the agent did and what you changed back. After a month it tells you which tasks have earned do alone and which still need you every time. It also gives you a clean record to work from if you later decide to automate more of the catalogue, a step worth planning carefully with this guide to ecommerce listing automation.

What does this template not cover?

These AI agent rules for sellers do not decide whether an agent may operate inside Seller Central or the Shopify admin at all, which depends on each platform's terms. They are not legal advice for your country, and they are not a full copy of Amazon's policies, which can change. The safest place to start is research, since OpenAI says "Proactive research itself doesn't send messages, change connected apps, or control a browser or computer." Begin with a week of research and audits before any row moves money.

ListingRVA AI sits on the draft side of the template: it generates marketplace-sized product images, Amazon A+ pages and listing copy that you review before anything goes live, and it does not connect to Meesho or Flipkart stores. Whatever writes your drafts, the rule stays the same. The agent prepares, and you publish.

Frequently asked questions

What are the most important AI agent rules for sellers?

Sort every task into do alone, ask first or never. Let the agent do audits, reports and drafts alone, make it ask before anything that changes what buyers see or moves money, and never let it offer review incentives, ask for review changes, send marketing messages, handle passwords or send spam.

Can an AI agent ask buyers to change a review?

No. An Amazon staff reply in the Seller Forums lists requesting changes to existing reviews, through Buyer-Seller Messaging, product inserts or external channels, as prohibited, and says Amazon has a zero-tolerance policy towards customer reviews manipulation. The same rules cover anyone associated with your account.

Should an agent be allowed to change my prices on its own?

Not at first. Our suggested default is ask first for every price, coupon and shipping change, with a floor and ceiling per SKU. An Amazon staff reply in the Seller Forums notes that Amazon regularly monitors prices, including shipping costs.

Is there a never option in OpenAI Custom Rules?

Not by that name. The closest behavior is "Ask you to take the action instead," which hands the task back to you. OpenAI also says rules do not grant access or override built-in safety requirements.

Should I give an AI agent my Seller Central or Shopify password?

No. Amazon's developer Acceptable Use Policy says never share access keys or passwords, an Amazon staff reply says not to add outside service providers as users, and Shopify's Terms say you are responsible for keeping your password secure.

Does pausing an agent stop everything it was doing?

No. OpenAI says pause stops the current main task but does not stop every delegated task or cancel future scheduled runs, and stopping work does not undo completed actions. Pause, cancel scheduled runs, then reverse changes by hand.

When can a task move from ask first to do alone?

Our suggested default is after a full week in which every instance of that task was correct with no edits from you. Keep a change log so the decision is based on what the agent actually did.

Published by Aashirvad Kumar, founder of ListingRVA AI. Drafted with AI assistance; the rules, figures and claims are checked against the sources linked in this article before publishing.