As long as the bias is not too strong, you are better off with a robust, outlier-insensitive estimate, than with an unbiased one. It would be interesting to do some analysis, to figure out the impact that a 10% bias has on your yield metric (measured as correctness of predictions, or revenue). The impact might be much smaller than 10%. Your model might be a bad model. It is better to reduce the variance generated by your model, rather than picking up a kosher (perfect) statistical estimate.

As a data scientist and ex-statistician, I violate these rules (especially #1 - #3) almost daily. Indeed, that's part of what makes data science different from statistical science. Some theoretical research should be performed about the maximum yield obtainable with non-kosher estimates. This article compares model-free confidence intervals with classic ones. The difference is very small even when the number of observations is as low as 50.