Learning error metrics

We're using a decision tree based classifier that assigns record to one of two classes (true /false). Model is producing answers with probability of this value. So to check how good are answers the squared error is used

(expected{0, 1} - result{0.0 .. 1.0})^2


This method gives quite good picture of learning curve, but the business problem is defined as find records that can be classified as "true" with highest probability. So from the business point of view even drastically biased results are ok, as long as records that are expected to be "true" gets even slightly higher ranks that those that are expected to be false.

My question is:

How to define metric to measure quality of this model from business point of view?