MMOOZZGG.com

How to build an AI agent

AI agent for marketplace sellers: what a seller can build alone

5 min read6 sources

An AI agent for marketplace sellers answers reviews and buyer questions, suggests what to add to product cards and tracks messages from the platform. 75% of buyers read reviews before a purchase, and about 40% expect a seller’s answer within a day. The agent keeps that deadline and hands hard cases to a person.

On this page · 5 sections

A developer is needed

What the agent does for a seller

A seller on a marketplace has four streams of text: reviews, buyer questions about a product, chat messages and notices from the platform itself. Each needs an answer by its own deadline, and each affects sales.

The agent writes answers from the product card, the instructions and the store’s rules. It works on the RAG principle: it finds the right passage in your documents and answers from it, with no model training.

The figures in the lead come from a survey of more than 1,000 buyers published in October 2024. Everyone who chooses a product sees the seller’s answer to a review, so it is written for future buyers, not only for the author of the review.

If you need a ready agent built for you, that is our service for Avito and marketplaces. Here we look at how it works and what you can build yourself.

Answers to reviews: built-in tools and your own agent

Platforms themselves give sellers AI for reviews. On 19 February 2026 Wildberries launched auto-replies: the seller chooses for which ratings the neural network writes and publishes an answer by itself, for example only for 4 and 5 stars.

Reviews with other ratings stay with the seller. Reviews where the product card lacks the information for an answer also wait for a person.

The built-in tool is enough for thanking buyers for high ratings. You need your own agent for what the platform does not know: your return rules, common problems with the product, the buyer’s history in other channels. The limits of the built-in tool are covered in the article on the Wildberries tool.

The agent prepares answers to negative reviews as drafts. A person publishes them until the log shows that the agent writes without mistakes.

Buyer questions and product cards

The agent closes a question about size, compatibility or contents from the product card and the manufacturer’s instructions. If the documents have no answer, it passes the question to a manager instead of inventing specifications.

Repeated questions show what the product card lacks. Once a week the agent collects them into a list: which specifications to add to the description so that buyers ask less.

Speed matters here too: about 40% of buyers expect a seller’s answer within a day. The agent answers without a queue, at weekends and at night, if it is connected to the platform through the platform’s official tools.

For each product category the agent gets rules: which promises it must not make (service life, compatibility the manufacturer has not confirmed), and when to offer a replacement or a return. The seller checks these rules on real questions from the past month before launch.

Platform messages and deadlines

Since 1 October 2026, Russia’s law on the platform economy (289-FZ) is in force. A platform must give a seller a way to exchange legally significant messages with it, and the exchange goes through the seller’s account or another way set in the contract.

For a seller, this means that a missed message in the account can cost a missed deadline. The agent reads the platform’s notices, picks out those with a deadline or a demand, and sets a task with a date for the person responsible.

The agent should not answer the platform in the store’s name on disputes and claims; that is the work of a manager or a lawyer. What exactly the law changed for sellers is covered in a separate article on the platform economy.

The agent puts tasks from platform messages where the team already sees them: in the CRM, a shared spreadsheet or a work chat. You do not need a separate program for this; it matters more to name a person responsible for each task.

Build it yourself or order it

You can build a helper that drafts answers yourself: you load the product cards and the store’s rules into an agent builder, copy a review and get an answer. This saves time on the text, but the copying stays manual.

An agent that collects reviews and questions from all platforms by itself, publishes answers and sets tasks in the CRM needs a developer: connections to several seller accounts, publishing rules by rating and topic, a log.

In reviews and chats, buyers leave names, phone numbers and addresses. This is personal data: under 152-FZ, Russia’s personal data law, it is stored in databases in Russia, and Roskomnadzor, the data protection regulator, is notified before any transfer abroad. More on the page about data security.

We launch such an agent as a pilot in 14 days, from $1,499. The metric in the contract: the share of reviews and questions answered within a day. What this looks like in numbers is shown in the analysis for a marketplace seller.

Questions and answers

Can you connect an agent to the seller account?

Yes, through the official tools the platform itself provides. Do not give the agent the login and password of your seller account: if it makes a mistake, the answer goes out in the store’s name, and the log will not show who approved it. A developer sets up the connection, and the agent gets narrow rights: reading reviews and questions and publishing answers.

Do you need an agent if the platform answers reviews itself?

Since February 2026, the built-in Wildberries tool answers reviews with the ratings you choose, for example 4 and 5 stars. Other reviews, buyer questions, platform messages and other marketplaces stay with you. Your own agent covers these streams and knows your store’s rules.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Forbes, 19 February 2026 (forbes.ru)
  2. New Retail, 18 October 2024 (new-retail.ru)
  3. ConsultantPlus, 289-FZ, Article 6, 31 July 2025 (consultant.ru)
  4. ConsultantPlus, 289-FZ, Article 23, 31 July 2025 (consultant.ru)
  5. arXiv 2005.11401, 22 May 2020 (arxiv.org)
  6. Brace Law Firm, 9 July 2023 (brace-lf.com)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us