An AI agent for marketplaces: product cards, reviews and replies

An AI agent for marketplaces is useful to a seller not to "run the store on its own" but to remove routine work: descriptions for hundreds of product cards, replies to repetitive reviews, analysis of return reasons. All these tasks are textual and repetitive, so they fit cheap models through an API well. Below: which tasks to hand over, how to build a pipeline with human review, how to calculate cost and what a model will not replace in analytics.

What to give the agent and what not

Tasks that fit have an input text or table and a clear result format:

Don't hand over without review: prices, discounts, delivery terms and warranties, any promises to the buyer, medical and legal statements. Marketplace rules for card and reply formatting change, so check current requirements in the marketplace's own documentation.

The best AI for marketplaces: how to choose

There is no universal "best" model, only a balance of price, context and quality on your texts. A practical approach:

1. Take 30–50 real examples (reviews, cards).
2. Run them through two or three models at different prices.
3. Ask a colleague to pick acceptable replies blind.
4. Compute the price for your volume and choose the cheapest of the acceptable ones.

Reference points from the catalog (prices as of October 18, 2026, $ per 1M tokens, input / output, strength index): deepseek-v4.1-flash 0.013 / 0.038 (39), glm-5.3-flash 0.04 / 0.04 (42), claude-sonnet-5-5 0.4 / 2 (48). More on choosing by index: model strength index↗.

Pipeline outline

A working variant for a small seller:

1. Export. You export reviews or characteristics from the marketplace seller account to CSV (see the marketplace help for how to export).
2. Processing. A script sends each row to the model with the rules from the system prompt.
3. Review. The result goes into a table with the status "under review": a human reads, edits and marks "ok".
4. Publishing. You publish approved texts yourself or by file upload, as the marketplace allows.

We don't recommend publishing without a human: a reply to a customer on your behalf should be reviewed, at least by sampling.

Example: review replies and classification

We ask the model to classify a review and prepare a draft in one request. Code on the OpenAI SDK; details are in the Python documentation↗:

import csv
import json
from openai import OpenAI

client = OpenAI(base_url="https://api.zukko.pro/v1", api_key="fetch_your_key")

SYSTEM = """You help a seller reply to customer reviews.
For a review return JSON: {"topic": "quality|delivery|size|packaging|price|other",
"sentiment": "positive|neutral|negative", "needs_human": true|false, "reply": "..."}.
Rules: polite, no excuses, do not promise a refund, discount or compensation,
do not invent facts about the product. If the review complains about a defect or health, set needs_human = true.
JSON only."""

def process(review: str, product: str, model: str):

The needs_human field solves the main problem: complex and complaint reviews automatically go to manual handling. For large volumes, send requests in parallel rather than one by one.

AI for marketplace analytics

The division of labor matters here. Code does the numbers: compute revenue, conversion, stock and trends from the table in pandas or Excel. The model can be given interpretation: group buyer comments by topic, briefly describe what changed, propose hypotheses that you then check against the data. If you hand a model a large table and ask it to "calculate", arithmetic errors are likely.

Cascade: cheap model first, strong one when needed

An economical scheme: a cheap model processes all reviews, and those where needs_human is true or the reply looks doubtful are resent to a stronger model, and only then to a human. The principle is described in the model cascade↗ article.

What it costs

Assumptions for the example: a review with instructions takes 800 input tokens, the reply 150 output tokens; 5,000 reviews a month. Total 4M input tokens and 0.75M output.

• deepseek-v4.1-flash: 4 × 0.013 + 0.75 × 0.038 = 0.052 + 0.0285 = $0.0805
• claude-sonnet-5-5 (OpenAI format, no cache discount): 4 × 0.4 + 0.75 × 2 = 1.6 + 1.5 = $3.10

Cascade: all 5,000 on deepseek-v4.1-flash ($0.0805), 500 hard ones again on claude-sonnet-5-5: 0.4 × 0.4 + 0.075 × 2 = 0.16 + 0.15 = $0.31. Total about $0.39.

Cards: 1,000 items, 1,200 input and 700 output tokens each, i.e. 1.2M and 0.7M.

• deepseek-v4.1-flash: 1.2 × 0.013 + 0.7 × 0.038 = 0.0156 + 0.0266 = $0.0422
• claude-sonnet-5-5: 1.2 × 0.4 + 0.7 × 2 = 0.48 + 1.4 = $1.88

This is price arithmetic, not a quality guarantee: whether a cheap model suffices for your tone and category is shown by a test on real examples. Fresh prices are always in the GET https://api.zukko.pro/v1/models response.

To try it, create a key in the dashboard↗ and run the script on 30 of your own reviews.

FAQ

Can the agent reply to marketplace reviews by itself?

Technically yes, if the marketplace provides an interface for it, but we advise leaving review to a human, at least by sampling. See the marketplace documentation for rules on automatic replies.

Is AI suitable for sales analytics?

For interpreting and grouping text, yes; for arithmetic, no: compute totals and trends in code, and ask the model to describe the result and suggest hypotheses.

Which model is needed for product card descriptions?

Start with a cheap one and compare it with a stronger one on 30–50 cards from your category. Choose the cheapest of those whose texts pass your review.

How do I avoid sending unnecessary data to the model?

Pass only the review text and the product name. Remove buyer names and contacts before sending.