From enterprise to marketplace automation
Я пришёл автоматизировать маркетплейсы из кровавого энтерпрайза
B · Useful, limited Field notes by the founder of Manaraga, an AI-agent vendor for fashion brands on Russian marketplaces, with specific pilot anecdotes and operating detail; all results are the vendor's own and largely unquantified beyond single examples.
Summary based on the full text·Original: about 12 minutes to read
In short
A Habr post (27 October 2026) by Ilya Letunov, founder of Manaraga, who builds AI agents for large brands and moved from insurance, banking and telecom work into fashion e-commerce on Wildberries, Ozon and Lamoda. His finding: marketplace data exists but brands cannot use it. Funnels live in Excel, alert bots get switched off, and managers can only watch a few flagship SKUs. The agent works when it learns each product's strategy from manager feedback and diagnoses causes per SKU; strategy itself stays with people.
Key numbers
- 18stalled SKUs in one pilot: 12 were priced above competitors, 6 had a main photo nobody clicked; cleared in two days
- 60-70%share of a category manager's time spent working out what happened yesterday, leaving 20-30% for action
- 30-60%share of clothing orders actually bought out on marketplaces; the rest is returned
- below 5%false-alert rate within the first week once managers can reject recommendations with a reason
- 10,000+SKUs at a large brand, each with about 30 metrics changing daily
As reported by the source.
Russian·Web page·opens habr.com
What it reports
- In his agents a router sends the simplest 50% of cases to a model and the rest to people; savings are usually 10-12%.
- Large brands chase turnover quotas with the marketplace (a 3 billion turnover target is the example), so they knowingly sell lines at a loss; a system that advises cutting traffic for negative unit economics looks "dumb" to them.
- Competitor price parsing is unreliable: "see also" prices are personalised per buyer. A T-shirt can range from 300 to 7,000 rubles, so brands track 3-4 direct competitors, not a market average.
- At 1 minute per SKU, a manager with 1,600 cards would need almost 27 hours a day. One brand has up to 10,000 colour models, about 60,000 barcodes.
- Marketplaces do not expose everything by API: Lamoda lacked funnel analytics until recently, Ozon gives organic data only on higher paid tiers, Wildberries order data is preliminary.
- Pilot: the default response to 18 stalled SKUs was a 10% price cut across the group. The agent found 12 overpriced and 6 with a weak main photo; targeted changes cleared stock in two days while keeping profit.
- Rule-based alerts (for example ad spend share above 15%) flood managers and get disabled. Feedback loop: a manager rejects a recommendation with a reason ("ads off on purpose, waiting for stock on 14 July"), the agent ties it to the SKUs and stays quiet until the condition changes.
- First failure: API numbers did not match the seller dashboard, so trust ended at the demo; they switched to dashboard report exports, at first downloaded by a staff member each morning under a read-only login.
- Second failure: asking clients to compute metrics and write strategy stalled a pilot for months; they now compute metrics themselves and use a 20-minute questionnaire. Back-calculating impressions from sales plans shows some plans are physically impossible.
- The system runs at 4 large manufacturers and suppliers, together "a few percent" of the Russian market.
How they got it
Pilot experience and client interviews quoted verbatim. No baseline, sample or measured uplift is reported beyond the single 18-SKU case and the under-5% false-alert figure.
Limits
- Vendor self-report with a promotional purpose; clients unnamed.
- The two-day result is one anecdote, without counterfactual.
- The 10-12% savings and the 50% routing share are stated as typical without supporting data.
- Specific to fashion on Russian marketplaces with turnover-quota contracts.
Why it matters for companies
Any firm selling through third-party marketplaces will recognise the pattern: data is available, but nobody can review thousands of SKUs daily, so most of the range runs on default rules. The useful design points are reconciliation with the numbers people already trust before any recommendation, per-SKU strategy captured through rejections with reasons, and diagnosis of cause (photo, size gaps, regional stock, reviews, weather) before reaching for a blanket discount. The author's conclusion that experienced managers make good decisions and lack coverage is a sensible framing for agent scope.
Key passages in translation
Russian original
У 12 артикулов цена реально проигрывала конкурентам. Но у 6 артикулов цена была в полном порядке! Проблема была в главном фото
For 12 SKUs the price really was losing to competitors. But for 6 SKUs the price was entirely fine. The problem was the main photo.
Russian original
Клиент смотрит в кабинет, агент смотрит в API, и числа по одному товару у них расходятся. Доверие к рекомендациям в такой ситуации кончается прямо на демонстрации.
The client looks at the seller dashboard, the agent looks at the API, and their numbers for the same product differ. In that situation trust in the recommendations ends right at the demo.
Russian original
решения у опытного менеджера хорошие и стратегию агенту отдавать нельзя. Проблема в охвате
An experienced manager's decisions are good, and strategy must not be handed to the agent. The problem is coverage.
Summary based on the full text. Read the original for the full argument and data.Read the original on habr.com