Three AI models instead of BI tables: how to analyze product profitability on Ozon

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

Why BI-tables Don't Solve the Problem

This is the trouble of almost every e-commerce seller with a large catalog. For 500–2000 product items, calculating the full P&L manually is hours per week, and in most cases, no one simply does it. A product is evaluated by turnover metric, but by real margin it can be in the red for years, completely unnoticed.

Analysis tools have long existed on the market. But almost all of them simply display lots of numbers in tables nicely. There's plenty of information, but no answer to the main question: what should I do with this now?

Classic BI-systems work by rigid rules and presets. They're not flexible, they don't look at the context of a specific product, they don't give advice.

Council of Three AI-Models

SKUmind solves this not with one model, but with a council of three different ones. There are different AI-models on the market: Claude, GPT, and others. Each is trained differently and looks at the task through different lenses. The idea is simple: give the same data to all three, and each will voice an opinion independently. Then a special arbitrator—also an AI—looks at where opinions agree, where they diverge, and chooses the most reliable conclusion.

Why exactly three? Because one model can start hallucinating and make up facts. A second can be too conservative and see risks that don't actually exist. A third can miss an important detail. When there are several, each checks the other and finds mistakes. It's like when a polyclinic assembles a medical council—one doctor sees one thing, a second sees another, a third notices what the first two missed. In the end, the diagnosis is more reliable than when one person looks at it.

Each of the three models analyzes per product:

• Real margin after all marketplace commissions and returns
• Effectiveness of each ruble spent on advertising
• Price potential: whether prices are overvalued or undervalued
• Seasonal trends and demand volatility over time
• Specific recommendations on what urgently needs to be changed

Implementation: API and Long Code Review

Under the hood, everything turned out to be much more complex than it initially seemed. Ozon doesn't publish all the needed metrics and numbers through the official API. We had to manually restore the logic from analysis of request traffic.

It helped that we could run two Claude sessions in parallel and give them to break down the same task from different angles, then compare results. After the application logic finally came together, the longest and most tedious part began: code review. Because this system directly impacts people's financial decisions, every line, every algorithm must be checked critically.

Now code review takes 60–70% of all development time. A long and exhausting process, but completely justified when it comes to other people's money.

What This Changes in the Industry

AI is moving from the category of experiments to the category of combat tools for work. Static BI-tables on rigid preset rules—that's yesterday. They're being replaced by models that can reason flexibly, like humans. For e-commerce this means: transparency in finances stops being a luxury for specialists, becomes an accessible standard for any seller with a large catalog.