Marketplace sellers: AI for listings, pricing and support

AI for marketplace sellers is not a single tool; it is four jobs that a seller on Amazon, Flipkart, Myntra and their own store does every day by hand, done from one system instead of four seller dashboards. Listing copy and attributes per marketplace. Price monitoring and repricing within limits. Review and question responses. Customer support that knows which order on which channel the customer means. This article explains what each job needs, why a catalogue of record is the foundation, where automation must stop, and what a build costs for a seller in India.

What AI for marketplace sellers means in practice

A seller with a few hundred SKUs across three marketplaces maintains thousands of listing variants, each with its own title rules, attribute schema and image requirements. Prices are checked against competitors by eye. Reviews and questions are answered when someone has time. Support arrives through each marketplace's messaging system plus WhatsApp. Marketplace automation replaces the copy-paste between dashboards with a system the seller owns: a catalogue of record with per-marketplace projections, a pricing engine with rules, and a support agent connected to every channel's orders. AI does the language work inside that system; the system does the plumbing. The retail industry page covers how this sits alongside what we build for brands with their own stores.

Seller dashboards vs a seller's own system

The catalogue of record

Everything depends on one clean product record per SKU: attributes, images, dimensions, compliance details, cost, and stock. Each marketplace gets a projection of that record, its own title format, its own attribute names, its own image specifications. When the record changes, every projection updates. Without this, listing optimisation AI produces better copy in four places that then diverge. Building the catalogue of record is the first phase of any seller system and usually the phase sellers underestimate, because it means reconciling the versions that already exist on each channel and deciding which is right.

Listing optimisation AI

Given the record and the marketplace's style rules, a language model drafts titles, bullets and descriptions that use the attributes shoppers search for, respect character limits and prohibited claims, and read naturally. The seller reviews drafts for the top SKUs and lets rules govern the tail. Two disciplines matter. First, the generated copy must not invent attributes: a material, a certification or a compatibility the record does not contain is a listing violation and a customer complaint. Generation is grounded in the record and checked against it. Second, changes are measured: a listing test on a set of SKUs, with search rank, click-through and conversion compared before and after, tells the seller whether the new copy works on that marketplace. Copy that reads better and converts worse is not an improvement.

Pricing: monitoring, rules, floors

Price monitoring collects competitor prices for the seller's SKUs from marketplace data where the terms allow it and from the seller's own tracked list elsewhere. Repricing then follows rules the seller writes: a floor at cost plus fees plus minimum margin, a ceiling, a target position against named competitors, and a rate limit on changes. A model can propose rules from history, but the rules themselves are visible and editable, and every price change is logged with its reason. Fully autonomous repricing without floors is how sellers discover a race to the bottom in a single afternoon. Margin after fees per channel is the number the pricing engine reports, because a sale on one marketplace at a price that looks fine can lose money after that marketplace's fee structure.

Review and question responses

Responses to reviews and product questions are a language task with a policy around it. The agent drafts an answer from the product record and the seller's return and warranty policy, in the tone the seller sets. Positive reviews and factual questions can be sent by rule; negative reviews, complaints and anything mentioning safety or a refund route to a person with a draft attached. The agent never writes a review, never solicits one in exchange for anything, and never posts as a customer; the Federal Trade Commission's rule on consumer reviews and testimonials, available at ftc.gov, sets out what regulators consider deceptive, and marketplaces enforce their own versions with account suspension.

Support across channels

A customer who bought on one marketplace and messages the seller on WhatsApp expects the seller to know. The support agent maps the customer to orders across every channel, answers order status from each marketplace's order data and the courier, and handles returns within the policy each marketplace allows, which differs by channel. Anything outside policy, or anything the marketplace requires the seller to handle through its own system, is routed accordingly. Multi-channel support is a customer service agent with more integrations; the design principles are in WhatsApp commerce with AI and returns and exchanges automation.

Where automation must stop

Three lines are firm. The system never fabricates product attributes or claims. It never manipulates reviews. And it never reprices below the seller's floor or outside a marketplace's pricing rules. A fourth is practical: marketplace APIs and feed formats change, and a system that assumes they will not breaks at the worst moment, so integration monitoring and a care plan are part of running it, not optional.

A worked example

A consumer electronics accessories seller with several hundred SKUs sold on two large Indian marketplaces, one international marketplace and its own store. Listings had been created at different times by different people and no two channels agreed on titles or attributes. Prices were checked against competitors weekly by an assistant. Support ran from four inboxes. The build began with the catalogue of record, reconciling the channel versions SKU by SKU, which took longer than expected and fixed several compliance gaps. Listing projections were generated per marketplace, tested on a subset, and rolled out. The pricing engine ran with the seller's floors and a rate limit, logging every change. Review responses were drafted for review, with positives auto-sent after a month. The support agent unified the inboxes and answered status and returns within each channel's policy. The assistant's job changed from copy-paste to exception handling and reading the margin-by-channel report.

Team and timeline

A seller system is a custom enterprise software build with AI components: a product engineer for the catalogue of record and interfaces, an integration engineer for marketplace APIs and feeds, and an AI engineer for listing generation, review drafting and the support agent, over eight to fourteen weeks in phases. The catalogue and listing phase falls under API development and integrations from $7,000 / ₹4.4L for the connectors, with the full system from $24,500 / ₹16L; the support agent adds a customer service agent scope from $12,500 / ₹8L. The Standard Care Plan at $2,500 / ₹1,60,000 per month covers marketplace API changes, which are frequent. All figures are on the pricing page. A ten-day Sprint Zero at $3,250 / ₹2,00,000 audits the catalogue and channel access and is credited to the build. The seller needs one owner for the catalogue and pricing rules, and access to each marketplace's developer programme.

Before you start: a checklist

• List every channel, the SKU count on each, and who currently edits listings
• Confirm API or feed access on each marketplace and the developer approvals needed
• Export current listings per channel and identify the SKUs where they disagree
• Write pricing floors, ceilings and rate limits as rules with an owner
• Decide which review and question types may be auto-sent and which need a person
• Collect return policies per channel, which differ
• Set the listing test: which SKUs, which metrics, how long
• Name the owner who reads the margin-by-channel report weekly

Glossary

• Catalogue of record: the single product record per SKU from which every marketplace listing is generated
• Projection: the version of a record formatted to one marketplace's rules
• Floor price: the lowest price the pricing engine may set, usually cost plus fees plus minimum margin
• Repricing: automatic price changes within rules in response to competitor prices or stock
• Listing test: comparing search rank, clicks and conversion before and after a listing change on a set of SKUs
• Margin after fees: revenue less marketplace commission, fulfilment and payment fees, per SKU per channel

Related reading

AI for D2C brands covers sellers who also run their own store; demand forecasting for retail inventory covers the stock side that multi-channel selling complicates; structured outputs and function calling explains how generated listings are kept inside a marketplace's schema.

Build the catalogue of record, put rules around pricing and reviews, and let AI do the language work inside a system you own rather than four dashboards you rent.