Will AI Replace Marketplace Managers: What's Already Automated and Where Humans Are EssentialArticle · digest for November 19
A seller at a wooden desk with a laptop and product samples next to an analytics interface under soft lamp light.
Article5 sections≈ 8 min

Will AI Replace Marketplace Managers: What's Already Automated and Where Humans Are Essential

"Who is actually going to be responsible if artificial intelligence makes a mistake and the store loses money?" Denis Vetrennikov, head of the SellerDen agency and creator of SellerDen AI, asked this question right at the start of the fifth episode of his podcast.

His perspective is both dual and honest. Every day, his agency sees what decisions sellers pay people to make, while his own software shifts part of the routine to algorithms. Debates about AI in e-commerce quickly split into two extremes: fear of sudden layoffs or skepticism after early clunky tests. To separate store economics from guesswork, let's break down the manager's full scope of work: product cards, reviews, analytics, pricing, and ad campaigns.

Why Fear of Replacement and Old Frustrations Slow Teams Down

The marketplace manager role is loosely defined. In one store, a person only manages product cards. In another, they run ads. In a third, they handle nearly the entire business. Every practitioner knows the typical weekly checklist: updating cards, setting prices and promos, running ads, tracking stock and shipments, analyzing sales and competitors, and answering reviews and customer questions.

That explains the wide income gap. According to Vetrennikov, pay ranges from 10–15 thousand to hundreds of thousands of rubles per month. Salaries depend on responsibility and impact on net profit, even though repetitive small tasks eat up a large share of the workday.

Two barriers stop teams from handing monotonous routine over to software. The first is fear of losing their job. SellerDen once had a manager who worried AI would replace them and insisted on doing everything by hand. Management pointed out where ChatGPT or specialized tools could speed things up, but the employee refused. Eventually they parted ways. Vetrennikov notes that other factors piled up too, yet rejecting new tools noticeably dragged down that person's speed.

The second barrier is memory of past failures. A couple of years ago, sellers and managers tried early versions of ChatGPT 4.0 or YandexGPT. They got 'plastic' photos, people with six fingers, and unreadable infographics that had to be redone ten times. After that, people closed the tab and went back to manual work. Over the past six to twelve months, generation quality has jumped sharply, so judging today's tools by old mistakes no longer makes sense.

Product Cards, Reviews, and Reports: What Gets Automated in Minutes

The gap between old expenses and current practice is clearest across four daily tasks. Some cut costs right away, while others save time, though human sign-off still matters.

Photos, infographics, and video. Launching a product line used to require a long chain: finding a photographer, designer, and copywriter, signing contracts, shipping samples, waiting for retouching, approving edits, and closing paperwork. With a large catalog, reshooting for every season and holiday cost a fortune. Today, AI builds visuals from source photos in minutes. By Vetrennikov's rough spoken estimate, generating one option costs under 10 rubles. That's a personal benchmark rather than a fixed rate, yet the contrast with studio shoots is obvious. Still, algorithms can distort product shape, add extra details, or ruin the layout, so a human must always do the final check.

SEO descriptions. Text models mastered Russian early on, making this task mature fastest. You upload specs and a raw manufacturer description, and you get keyword-rich copy tailored to different audience segments, seasons, and sales. In this setup, the manager sets the task and verifies product facts.

Review replies. This is an obvious area for savings, especially when order volume grows and the owner wonders whether to hire part-time or full-time staff. Even with few reviews, software takes off unnecessary load. Vetrennikov points out a basic rule: a seller replies in the thread mainly for future shoppers who scroll through cards to judge if the store is reliable. Algorithms handle 4–5 star ratings effortlessly. Negative reviews of 1–3 stars are trickier. You have to train the model on your brand tone and check its drafts so a clumsy template doesn't scare off the next buyer.

Card audits and basic analytics. Tools already check cover images and text, spot weak points, and break down sales, expenses, and inventory. However, reports depend on complete source data. Without real cost of goods, marketplace fees, logistics, and accurate stats, AI will produce a smooth summary without reliable net profit.

Advisor or Driver: Where the Red Line Lies

Debates around automation constantly mix up two different things: an AI assistant and an autonomous AI agent. An assistant reads the numbers, spots drops, and suggests next steps, but it never clicks buttons in your seller dashboard. An autonomous agent sets its own tasks and changes store settings.

A real-world benchmark on the market is the 'Smart Assistant', which Ozon began rolling out in test mode in late April 2026 for 'Administrator' and 'Manager' roles. Powered by the Qwen 3.5 language model, it keeps context across follow-up questions. The tool summarizes yesterday's metrics, explains calculations like 'Ordered amount', suggests tips on shipments and cards, and pulls answers from the Ozon Knowledge Base. According to Ozon, the assistant can resolve up to 20% of routine seller support tickets. It cannot change prices or ad bids—it's a digital advisor without access to manual controls.

An autonomous AI manager goes much further. The software has access to sales, stock, unit costs, and ads. It spots an order drop, finds the cause, adjusts a price or bid, and measures the result. Sellers aren't ready to hand over direct control of their price list and ad budget yet, because the financial risk is too high.

Who Pays for Failures and How to Build Guardrails

The reason for caution is simple: software doesn't pay out of pocket. If an algorithm drops a price below cost or burns your budget overnight on junk search terms, the owner or responsible manager covers the loss. There are no industry-wide stats on how often this happens, yet every seller who has ever cleaned up after an accidental discount knows the risk.

Why You Shouldn't Let a Cheap Model Touch Prices or Ads

An AI that writes decent descriptions and replies to five-star reviews can fail badly at unit economics. Without factoring in fees, return logistics, and margins, an algorithm can easily try to boost orders with a money-losing price. Your store's bank account will pay for that experiment.

Giving automation access to money is only safe behind guardrails and manual approval—strict boundaries the system cannot cross without your confirmation. Vetrennikov highlights four rules:

  1. Set hard limits: a minimum price floor that covers marketplace deductions, and a daily cap on ad spend.
  2. Require manual approval for any edits outside those limits, and at launch, for every critical action.
  3. Match the model to the task: numbers, finance, and analytics require advanced or specialized models instead of cheap basic ones.
  4. Early on, check both the final recommendation and the model's step-by-step logic while it learns your store's rules.

Four areas remain completely human. Software won't pick your business model or strategy—it gathers niche data, but the entrepreneur makes the call. It won't negotiate with suppliers, though it can draft emails and compare price lists. It cannot manage financial risks without a safety net, since it can misread a formula or miss an expense line. Finally, AI gets lost in emergencies: a lost shipment, a marketplace dispute, an angry customer conflict, or a sudden contract change all demand a human response.

Area of WorkWhat AI HandlesWhat Stays Human
Photos, infographics, videoCard visuals from source imagesProduct accuracy and layout check
SEO copy and auditsKeywords, seasonal updates, spotting weak spotsSetting tasks and verifying facts
Reviews and questionsReplies to 4–5 stars and draftsBrand tone and 1–3 star negatives
Store analyticsBreakdown of sales, stock, and costsFull unit costs and verifying insights
Pricing and adsSpotting drops and edits within limitsGuardrails, approval, and model choice
Strategy, suppliers, crisesGathering inputs and drafting emailsNegotiations, risks, disputes, emergencies

AI Agent Operator: How the Role Is Changing and Who to Hire

Churning out descriptions, covers, and standard replies by hand is moving to software. Meanwhile, financial analytics, budget control, assortment planning, supply management, and handling unusual problems are gaining weight. Vetrennikov predicts the rise of an 'AI agent operator for marketplaces'—one specialist with the right toolset doing the workload that used to require several employees. Hiding from automation makes no sense. The manager who learns these tools, tracks updates, and grows net profit wins.

If you are just starting on marketplaces, it's easier to handle part of the daily operations yourself with AI. Cards and review replies take far less time, so you don't have to hire staff right away. If you already run a medium or large business, hire managers who have real hands-on experience with AI and a habit of updating their toolkit, because software capabilities shift in just a few months.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 4