Amazon Buyer-Seller Messaging: AI Replies Within the Rules

Let AI write the first draft, then screen it against Amazon's Communication program policy before sending. Keep replies about the order, with no external links, promotion, contact details, emojis or review requests, and answer within 24 hours. Amazon holds you responsible for every message, so a person approves each draft.

Most sellers who get into trouble over messaging did not set out to break anything. A friendly sign-off with the shop's website, a "thanks for buying, a five-star rating would help us" line, an emoji: each looks harmless, and each is on the list of things the policy restricts. AI is very good at writing exactly that warm, slightly promotional text, which is why it needs a rulebook around it.

Follow me on Instagram@sagnikteaches

Why a fluent draft is the risk, not the saving

A chatbot writes in the register of a good customer-service email from any shop on the internet. On a retail website that includes an invitation to follow the brand, a discount code for next time and a polite request for feedback. Inside Amazon's buyer-seller messaging, most of those are exactly what Amazon forbids. So the saving is real (a draft in seconds rather than three minutes) but only if the draft is edited for the platform.

Connect on LinkedInSagnik Bhattacharya

The rules below come from Amazon's Communication program policy page in Seller Central help (many guides still call it the Communication Guidelines), read on 5 November 2026. Amazon edits that page from time to time, so open it in your own account before you save templates, and treat this tutorial as the shape of the rules rather than the final word.

Subscribe on YouTube@sagnikteaches

What you may send, and what Amazon blocks

Amazon draws a line between messages needed to complete an order or answer a buyer, and everything else. Replies to a question the buyer asked are the safest ground. Proactive messages (ones you start) are narrow: resolving a fulfilment problem such as a delay, asking for a detail needed to finish the order (an engraving line, say), questions needed to complete a return, sending an invoice, scheduling delivery of a heavy or bulky item, and one neutral review request. Each must go out within 30 days of the order completing, carry the full 17-digit order ID and be written in the buyer's preferred language.

On links, Amazon's wording is precise: external links are prohibited unless they are secure (https) working links needed to complete the order, or links to Amazon. Links to warranty or registration sites that are not already in your listing, and links to external videos, are named as prohibited too. Several seller-tool guides simplify this to "no external links at all". For AI drafts I follow that stricter reading, because leaving a link out costs nothing and a blocked message costs you the reply window.

Two further points catch people. Amazon says it may block or modify any message that contains prohibited content, and breaches can lead to limits on your proactive messages or action against your selling privileges, so prevention matters more than argument. And review wording is policed hard: no asking only happy buyers for reviews, no offering anything in return, no asking a buyer to remove or change a review, and no "contact us first if anything went wrong, otherwise please leave a review" line, which the policy names specifically. That last sentence is exactly what a chatbot writes by default. For what happens when a conversation escalates, the A-to-z claim evidence tutorial picks up where this one stops.

The 24-hour clock

Amazon's policy says sellers are expected to reply to buyer messages within 24 hours of receipt, and it says nothing about pausing the clock for weekends, so plan cover for them. After 48 hours, a buyer on a seller-fulfilled order may become eligible to file an A-to-z claim. For seller-fulfilled orders, the Feedback Manager page also shows your average contact response time over the past four weeks. A message that genuinely needs no reply (a buyer's "thanks, all sorted") can be marked No Response Needed, which takes it out of the response-time calculation. Do not use that button on a real question to protect the figure: it hides a customer problem that tends to come back as a return or a claim.

One more split saves time. If the order was fulfilled by Amazon, Amazon's own messaging FAQ tells sellers to direct the buyer to Amazon customer service, because Amazon handles delivery, returns and refunds on those orders. A short reply that says so, with nothing else, is the right draft.

AI helps here in one specific way. The slow part of a reply is not typing, it is finding the facts: which size is the fitting, is the glass tempered, what does the warranty cover. If you keep a short fact sheet per product and give it to the AI with each question, the draft arrives with the right answer in it. That sheet is worth building more than any prompt wording.

Build the product fact sheet first

An AI draft is only as accurate as what you give it. Without facts it will guess, and a confident wrong answer ("yes, it fits a 2019 model") becomes a return, an A-to-z claim or a negative review. So begin with a one-page sheet for each of your top ten products.

Say a small furniture maker sells a flat-pack oak bookcase and ships the orders itself. The sheet might read as follows (illustrative).

PRODUCT: Oak bookcase, 5 shelves
Dimensions assembled: 180 cm high, 80 cm wide, 30 cm deep
Weight: 28 kg, ships in two boxes
Shelf load: 20 kg per shelf, evenly spread
Assembly: about 45 minutes, two adults, Allen key included, wall bracket included
Finish: oiled oak veneer; wipe with a damp cloth, no polish
Not included: wall plugs (depend on wall type)
Warranty: 2 years on defects in materials or manufacture
Returns: per Amazon's standard return window for the item
Do not say: anything about suitability for load-bearing walls, fire safety, or children climbing

The final line matters. It lists the claims the AI must never make for you, which is where safety and liability sit. Most buyer questions are answered by a line in this sheet; the rest go to a person.

A prompt that drafts inside the rules

Put the rules into the prompt itself, so the draft starts closer to compliant. Save this as a note and paste it each time, followed by the product sheet and the buyer's question with identifying details removed.

You are drafting a reply to a buyer message on Amazon for a small seller.
Use ONLY the product facts I give you. If the facts do not answer the question, say so in a line after the draft starting "MISSING:" and do not guess.

Rules for the reply:
- Plain text only. No emojis, no bold, no links of any kind, no website name, no email address, no phone number.
- Answer what the buyer asked, in the order they asked it. No promotion, no mention of other products, no discount.
- Do not ask for a review, a rating or feedback, and do not hint at one.
- Do not end with a standalone thank-you; start with the answer.
- If a refund, return or replacement is relevant, say what the seller will do and what the buyer needs to do inside Amazon's returns process.
- Keep it under 120 words. Use "[first name]" and "[your name]" as placeholders.

PRODUCT FACTS:
[paste sheet]

BUYER MESSAGE (personal details removed):
[paste]

Notice the "MISSING" instruction. It turns a silent guess into a visible gap, which you then fill from your own knowledge. It is the single most useful line in the prompt.

A sample draft and what to fix in it

Take the bookcase seller and a buyer who asks: "Can this hold my record collection? About 25 records per shelf. And does it come with wall fixings?" Here is the kind of reply a model returns (illustrative output).

Hi [first name],

Each shelf is rated for 20 kg spread evenly. A record weighs roughly 180 to 220 g, so 25 per shelf is around 5 kg and sits comfortably inside that.

A wall bracket is included, but wall plugs are not, because the right type depends on your wall. Your local hardware shop can advise.

Hope this helps! Feel free to browse our other furniture on our store page. Leaving a review would mean the world to us.

Kind regards,
[your name]

The first two paragraphs are good. The third contains three faults, none obvious at a glance. "Browse our other furniture on our store page" is a promotion and a pointer away from the message. The review line is a prohibited request. And "Hope this helps!" is harmless alone but, with the rest, tips the message into chatty filler. Also look at the record weight: the model supplied "180 to 220 g" from general knowledge, not from my sheet. It is plausible but unverified, so I would keep the shelf rating, soften the per-record estimate to "about", and let the buyer judge. The fixed version ends at "your wall" with a clean sign-off.

The screening pass: a checklist before sending

Do not trust the model to police itself. Even with the rules in the prompt, a draft can slip, so run a short screen. You can do it by eye in ten seconds, or ask a second AI pass to flag, then confirm yourself.

1. Does the first sentence answer the question the buyer asked?
2. Any link, domain name, "@" sign or phone number? Delete it.
3. Any mention of another product, a discount, a "next time", or a store page?
4. Any request, hint or nudge about reviews, ratings or feedback?
5. Any emoji, exclamation-heavy tone, or decorative formatting?
6. Does every factual claim appear on the product sheet? If not, verify it or cut it.
7. Does it promise something you will not do, such as a refund outside policy or a delivery date you do not control?
8. Is it free of anything about suitability for safety-critical use?

A second-pass prompt makes this faster for long messages:

Audit the draft below against these rules and list every breach as a quoted phrase with the rule it breaks. Rules: no links or domains; no email or phone; no promotion or mention of other products; no review, rating or feedback request or hint; no emojis; no claim not supported by the product facts; no standalone thank-you. If there are no breaches, say "No breaches found." Do not rewrite the draft.

PRODUCT FACTS: [paste]
DRAFT: [paste]

The "do not rewrite" line stops the model from fixing silently, which would hide what it changed. You want to see the breaches and decide.

Three messages, three different handlings

A simple fit question

A cleaning company sells a refill pack of microfibre cloths and a buyer asks whether they are safe on a glass hob. This is the easy case: one fact, one answer. If the pack's care label says "do not use with abrasive cleaners" and nothing about glass, say what the label says and stop. Do not stretch it to "yes, perfectly safe" because that sounds helpful. The AI will want to; the fact sheet should say "no claim about hob surfaces".

A complaint with some anger in it

A pest-control firm sells ultrasonic repeller plugs. A buyer writes: "Useless. Mice are still in the kitchen. I want my money back." The temptation is a long, defensive reply. A short one works better: acknowledge the problem, say how to start a return inside Amazon, and state plainly that the product does not guarantee results in every home, if that is what your listing says. Leave out anything about the buyer's review. A prompt line such as "acknowledge the problem in one sentence, no excuses, no blame" gives a more measured draft than the model's default.

A question with a hidden safety angle

A car repair garage sells a dash-cam hardwire kit. The buyer asks if it is safe to wire it straight into a fuse box themselves. The right reply points to the manual and to a qualified fitter, and says the seller cannot advise on a specific vehicle. This is where the "Do not say" line in your fact sheet earns its keep. If the draft gives step-by-step wiring advice, it fails the screen even though it is helpful and compliant in tone.

When the reply is not the AI's to write

Some messages should reach a person before any draft exists: anything mentioning injury or damage to property, an A-to-z claim, a chargeback, a legal threat, a counterfeit or "not as described" accusation, or a message that looks like a competitor probing for supplier details. A simple rule: words such as "injured", "lawyer", "fire", "burn", "sue", "fake", "counterfeit" or "A-to-z" route to you, not to the AI.

If you have more than a handful of messages a day, add a triage step. Ask the AI to sort each incoming message into "routine, answerable from the sheet", "needs the owner" or "might be spam or off-platform contact". Even a rough sort saves reading time, and the "needs the owner" pile is the one that protects the account.

A filled-in template library, and why it needs a label

Once a type of question recurs, a saved reply is faster than drafting. It is worth listing every message a buyer can receive from you (manual replies, saved templates, automated sequences, package inserts and anything sent by connected software) and re-auditing that list each quarter. Build the library from replies that have already passed your screen. For proactive messages, Seller Central's Contact Buyer page also offers templates with Amazon-supplied wording, translated automatically into the buyer's store language; where one fits, use it instead of a custom draft.

A small set for the bookcase seller (illustrative):

• Missing customisation detail: "Hello [first name], to finish order [17-digit order ID] I need the engraving text you would like, up to 20 characters. Please reply with it as written, including capitals. [your name]"
• Delivery scheduling: "Hello [first name], for order [17-digit order ID] the carrier needs a delivery window for this two-box item. Which of weekday morning or weekday afternoon suits you? [your name]"
• Damaged on arrival: "Hello [first name], I am sorry the bookcase arrived damaged. Please start a return or replacement request from your orders page and attach a photo of the damage and the outer box. I will approve it as soon as it arrives. [your name]"

Put the date of the last audit and the owner's initials in a spreadsheet next to each template. A template nobody has read in a year is where blocked messages come from.

Where connected tools and automation fit

Amazon now runs its own Seller Messaging Assistant, a conversational assistant that answers common buyer questions (delivery status, cancellations, returns, invoices, product details and warranty) from Amazon's data and the auto-responses you configure, and passes the rest to you. Check what it is set to say before you build your own drafts for the same questions; your AI drafts are mainly for what it escalates.

Many sellers also connect helpdesk or review-request software to Buyer-Seller Messaging. Whatever sends the message, you remain responsible for it. Before connecting a tool, confirm that any review request goes through Amazon's Request a Review feature, not a message the tool composes, and that its templates contain none of the restricted elements: links, promotion, emojis, tracking pixels. If a tool offers "AI-written replies", ask whether a person approves each message and what data goes to the AI provider. For an overview of what Seller Central already includes, see the Seller Central AI features tutorial; for whether to hand message handling to a person instead, the virtual assistant comparison helps.

Privacy: what goes into the chatbot

A buyer message can contain a name, delivery address and sometimes a phone number. Remove them before pasting. The draft does not need them, and you can add "[first name]" back in yourself. Check your plan's data settings: consumer and business plans treat your content differently, and the setting names change often, so read the provider's current data-use page rather than relying on memory. Amazon's Seller Code of Conduct adds a harder rule: customer information you receive to fulfil orders may be used only to fulfil those orders and may not be shared with any third party. A chatbot provider is a third party, so the name and address never go in.

A short sum: is the setup worth it?

Say a seller gets 18 buyer messages a day and each takes four minutes to find the facts and write. That is 72 minutes a day. If the fact sheet and prompt cut a routine message to about 1.5 minutes and 12 of the 18 are routine, the day drops by roughly 30 minutes (12 messages saving 2.5 minutes each). Over a month of 26 working days that is around 13 hours. It is an illustration with my assumptions, not a measurement; time your own messages for a week before you decide. The setup (fact sheets for ten products, a saved prompt, a checklist) takes perhaps three or four hours once.

How to check the routine is working

After two weeks, pull your message log and look at four things. Were any messages blocked or flagged by Amazon? What is your response-time figure? How many drafts needed more than a light edit? And did any reply lead to a return or complaint that a correct answer would have avoided? If blocked messages appear, look at the wording of those exact messages first; usually one phrase repeats. Fold that phrase into the prompt's rules and the checklist. Fix the system, not just the single message.

A weekly look at your reports also catches patterns no single reply shows. A rise in the same question suggests a listing gap; the review analysis tutorial shows how to turn recurring buyer language into bullet and image changes, and the weekly reports review gives a rhythm for it.

Slips that began with a helpful draft

One typical failure (an illustration): a garage selling brake-cleaner spray let a model reply to a question about using it near a hot engine. The draft said it was "fine in normal use". The product page said it was flammable. The seller caught it on the read-through and replaced the line with a pointer to the label's safety wording, which is the whole reason for the human step. Another: a seller kept a saved template ending "Visit our website for care tips." It had gone out for months before a blocked message revealed it. The cure was the template audit above.

A third: pasting the whole buyer message, delivery address included, into a chatbot "to give it more context". The address added nothing to the answer. Stripping it is now a habit.

What stays human

Decide in advance which decisions the AI never makes: refunds beyond policy, anything about safety, anything about legal claims, and every case where the buyer is upset enough to leave a review you would like to talk them out of (you must not try). The AI is a fast first draft for routine messages. The fact sheet, the rule list and the ten-second read are what keep the account safe.

If you would rather map this with someone, that is what my AI implementation consultation covers; see the contact page.

Questions sellers ask about AI and buyer messages

Can I paste a buyer's message into ChatGPT or Claude?

You can if you strip the name, address, phone number and order details first. Paste only the question itself, plus the product facts you hold. The buyer's identity is not needed to draft a good reply, and removing it keeps personal data out of a tool you may not control.

Can AI answer buyers automatically on its own?

Amazon's own Seller Messaging Assistant already answers some routine buyer questions automatically, and connected tools can send through Buyer-Seller Messaging, but I would not let a custom AI draft go out unread. Amazon holds you responsible for every message, whoever or whatever wrote it. Ten seconds of reading per reply is the cheap safeguard.

Is it fine to ask a happy buyer for a review in my reply?

Keep it out of your reply. Amazon's policy points sellers to its own Request a Review feature, which sends a neutral request, and allows one request per customer per order. Hinting that a good rating would help, or asking only satisfied buyers, is wording the policy prohibits.

What if the AI reply is helpful but includes my website?

Delete the link. Amazon's policy prohibits external links unless a secure link is needed to complete the order or points to Amazon, and a link to your own site is neither. Put the instructions in the message itself, or make sure they are already on the product page.

How often should I re-check my templates?

Re-read the current Communication program policy in Seller Central help whenever a message is blocked, and at least once a quarter. Rules shift, and a saved template written a year ago may now contain wording Amazon no longer accepts.

Further reads

• How Small Amazon Sellers Use AI for Listings and Reviews — A wider view of where AI fits in a small seller's week.
• Amazon Seller Assistant Explained: Check Its Advice First — Check Amazon's own assistant's advice before acting on it.
• AI Tool for Amazon Listing Copy: Bullets, Search Terms and Checks — Fix listing gaps that keep generating the same question.
• Amazon FBA Reimbursement: Spot Lost and Damaged Stock Using AI — Recover money for stock that never reached a buyer.
• How to Answer Etsy Customer Messages Faster With AI — Compare marketplace messaging rules on a different platform.
• AI for TikTok Shop Sellers: Listings, Creator Samples and Replies — Turn verified product details into listing drafts, sample outreach and buyer replies without letting AI invent claims.
• AI for WooCommerce Stores: Descriptions and Support Replies — Draft descriptions in a controlled CSV batch and prepare support replies from verified order details.
• Amazon A+ Content With AI: The Built-In Generator and Checks — How to use the AI Ready modules in A+ Content Manager, and a pre-submission review for claims, image text and facts.
• Amazon Plan of Action: Draft an Appeal With AI, Truthfully — Build a fact file first, then let AI draft the root cause, corrective actions and preventive actions strictly from your own records.
• Does Amazon Allow AI-Generated Product Images? What Passes Review — Amazon's image guide does not ban AI images, but it does ban inaccurate ones. A rejection checklist built around one listing.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Amazon Seller Central help pages read 5 November 2026: Communication program policy (G1701), Selling policies and seller code of conduct (G1801), Buyer-Seller Messages FAQ, Seller Messaging Assistant overview and Customer Service Quality Insights; confirm the live wording in your own account.