An AI pricing engine lifted profit and sales at the same time
A wholesaler competing in low-margin, sealed-bid sales priced largely by hand. We built an AI pricing engine from scratch - and after the first version fell flat, the second raised profit 15–20% and sales over 10% together.
The challenge
Our client competes in low-margin, sealed-bid sales where the lowest price wins - but no one ever sees what rivals bid, and order quantities are unknown in advance.
Bidding blind, every day
The business runs on sealed bids where the lowest price wins and competitors' offers are never revealed - like a poker game where you can't see the other players' cards or bets, only whether you won or lost. Order quantities are unknown up front, which makes margins hard to forecast.
Win rate vs. profitability, with no clear dial
Every price was a trade-off between winning the deal and protecting the margin, with no reliable way to balance the two. Too low and profit leaked; too high and the sale went to a competitor.
Pricing was manual and inconsistent
Prices were set largely by hand, leaning on estimated minimum margins and maximum prices. The result was inconsistent outcomes across the catalog and significant ongoing pricing effort.
What we built
We built an AI-driven pricing engine from scratch - data analysis, feature engineering, model development and production deployment - in under six months. It sets optimal daily prices, and the breakthrough came from not giving up when the first version underdelivered.
- 01Modeling
Simulate the bidding, predict win probability
The model simulates bidding and order behavior and predicts the probability of winning at different price points, then optimizes each price to maximize expected profit while staying competitive. Prices adapt to demand patterns and inventory levels.
- 02Setback
The first model just copied the old prices
Trained on history alone, the pure machine-learning model replicated the exact pricing the client had always used - an expensive way to do what they already did. It could predict from past bids but couldn't optimize: it looked backwards instead of learning from its own results.
- 03Breakthrough
A tuning layer that learns - and wins
We added a rule-based layer on top: it compares results against the old approach in real time, lowers prices automatically when losses build up, and gradually learns which products can sustain higher margins. Simple, but it changed everything.
Results
The second iteration delivered what usually can't be had together: more profit and more sales, at the same time.
AI-optimized pricing raised profitability by 15–20% - lowering prices where money was being lost unnecessarily and raising them where margin was being left on the table.
Sales volume grew by over 10% at the same time - two metrics that normally trade off against each other.
Data analysis, feature engineering, model development and deployment, delivered in under six months with a short payback period.
Optimal prices are now set automatically every day, cutting manual pricing work and making outcomes consistent across the whole catalog. The solution keeps generating value month after month.
Beyond the numbers
- Profit and sales usually move in opposite directions - here both rose together
- The real lesson wasn't about algorithms: it was about trying again after the first version didn't work
