Three AI Models Instead of BI Spreadsheets: How to Analyze Product Profit on Ozon

An Ozon seller sees that a product is at the top: turnover is growing, the catalog is thriving. But at the end of the month the money somehow doesn't arrive. Not because there is no turnover at all, but because it's unclear where the money went after marketplace commissions, returns and advertising spend.

Why BI spreadsheets don't solve the problem

This is a problem for almost every e-commerce seller with a large catalog. With 500–2000 product listings, calculating a full P&L by hand takes hours every week, and in most cases nobody does it. A product can look good by turnover while sitting at a loss on real margin for years, completely unnoticed.

Analytics tools have been on the market for a long time. But almost all of them simply display lots of numbers in tables, nicely. There is plenty of information, but no answer to the main question: what should I do with this now? Classic BI systems run on rigid rules and presets. They are inflexible, they don't look at the context of a specific product, and they give no advice.

A panel of three AI models

SKUmind tackles this not with one model but with a panel of three different ones. The market has various AI models: Claude, GPT and others. Each is trained differently and looks at a task through a different lens. The idea is simple: give the same data to all three, and each offers its opinion independently. Then a dedicated arbitrator, also an AI, checks where the opinions agree and where they diverge, and picks the most reliable conclusion.

Why three? Because one model may start to hallucinate and invent facts. A second may be too conservative and see risks that don't actually exist. A third may miss an important detail. When there are several, each checks the others and catches errors. It's like a medical panel at a clinic: one doctor sees one thing, a second sees another, and a third notices what the first two missed. The resulting diagnosis is more reliable than when a single person examines the patient.

For each product, each of the three models analyzes:

• Real margin after all marketplace commissions and returns
• The efficiency of every ruble spent on advertising
• Pricing potential: whether prices are overvalued or undervalued
• Seasonal trends and demand volatility over time
• Specific recommendations on what urgently needs to change

Implementation: the API and a long code review

Under the hood, everything turned out to be much harder than it first seemed. Ozon does not publish all the necessary metrics and figures through its official API. The team had to reconstruct the logic manually by analyzing request traffic. It helped that two Claude sessions could be run in parallel, each given the same task to examine from a different angle, and then their results compared.

Once the application logic finally came together, the longest and most tedious part began: code review. Because this system directly affects people's financial decisions, every line and every algorithm must be checked critically. Code review now takes 60–70% of total development time. It is a long and exhausting process, but entirely justified when other people's money is involved.

What this changes for the industry

AI is moving out of the experimental category and into the category of working tools. Static BI spreadsheets built on rigid preset rules are yesterday's approach. They are being replaced by models that can reason flexibly, like a person. For e-commerce this means financial transparency stops being a luxury for specialists and becomes an accessible standard for any seller with a large catalog.