What is being optimized
Price optimization picks the price that maximizes expected margin: the margin at a price multiplied by the probability of winning at that price. The machine learning part is estimating that probability, the win-rate curve, from your history. Everything else is arithmetic and policy.
The win-rate curve, with an example
For one segment and part family, the model learns how win rate falls as the price rises from list minus 8% to list plus 3%. Here is the shape it found for workshops buying deep-groove bearings at one distributor, and the price it picked.
Why per segment
The curve above is useless for fleets, who in the same data paid up to 4% over list when stock was short, and for OEM accounts on contract, who do not respond to price at all. The model learns one curve per segment and part family, with wider bands where history is thin. Segments come from your ERP's customer fields, and customer intelligence shows the learned tolerance per account.
From curve to a price a rep can send
The optimum on the curve is a candidate. The policy floor, ceiling and rounding apply. Stock pressure and competitor position shift it. The confidence decides whether the engine moves at all. The result is written to the quote line with its factors, in about a second. Reps at Mineral Circles Bearings send 97% of these prices unchanged.