Imagine playing poker where you can't see your opponents, don't know their cards, and only find out whether you won or lost, never by how much.

This was precisely the challenge facing a B2B company that sells its products through sealed-bid auctions: the lowest price wins, but competitors' bids remain secret. The goal was clear but demanding: optimize pricing to win auctions with the best possible margin.

The result? A project that started with failure, turned into success through iteration, and increased profitability by nearly 20%.

‍

‍

The Challenge: Pricing in the Dark

In a closed auction, pricing is a constant balancing act.
The difficulty was compounded by imperfect data: only the winning bids were known, no one knew how much was lost bids. In other words, we were flying blind.

‍

‍

“Failure” and Why It Was Necessary

Our first approach was to build a CatBoost-model to simulate auction behavior. Technically, it worked, it replicated historical data perfectly, but it didn’t improve results. We had created a sophisticated way to repeat what the client was already doing.

In many AI projects, this would be the point where panic sets in. But in data science, a failed test is just another datapoint. We realized that modeling probability of sale wasn’t enough, we needed an active optimization layer to push prices intelligently without hurting conversion.

‍

‍

The Turning Point: The Tuning Layer

In the second iteration, we didn’t start over, we built on top of the existing model.
We introduced a heuristic tuning layer that dynamically adjusted AI-generated prices:

• Performance comparison: The model’s pricing was continuously compared against the client’s old system to track uplift.
• Dynamic adjustment: If performance dropped, prices were automatically tuned down to regain volume.
• Flexible floor: A dynamic minimum price ensured slow-moving products could be discounted while high-demand items could be priced higher.

‍

‍

The Results: Breaking the Classic Trade-Off

The second pilot hit the mark. Traditionally, you have to choose between margin and volume. This time, both improved.

Profitability rose by 15–20% across tested contracts, and sales volume increased by almost 10%.
With a project cost of roughly €30,000–40,000, the ROI was exceptional, creating a recurring revenue uplift that paid itself back quickly.

How? The AI identified products that were historically overpriced and slightly lowered them to win auctions and simultaneously raised prices on underpriced items without losing bids.

‍

‍

Lesson: Iterate or Die

The biggest takeaway wasn’t about algorithms, it was about persistence.
If we had stopped after the first pilot, this would have been a failed project. Instead, we treated it as a learning step, iterated, and unlocked substantial business value.

Three key lessons for your next AI project:

1. Understanding data takes time. Real data is messy. We spent two weeks just cleaning and structuring it.
2. Don’t be afraid to pivot. The first model rarely works perfectly. That’s normal: value emerges in iteration.
3. Solve the business problem, not the math problem. The elegant CatBoost model didn’t save the project, the heuristic business logic layer built on top did.

‍