Which of the following are true about your data organization: We have so much data we can train accurate deep nets for every question we care about. Our problems are so well specified that we just iterate on prediction accuracy and make use of new data as it becomes available. We're already collecting every kind of data we ever could, so we only try to solve questions that data can answer. If any of the above apply to you, what are you doing here? Go back to your desk and fix the world! For the rest of us, data science involves some modeling, and a lot of negotiation to make sure the data we're capturing, the questions we're asking, and the value we're trying to produce for customers all line up. That negotiation is the core of the interaction between data engineers, data scientists, and product managers. In this talk, I'm going to discuss that negotiation in the geekiest way possible: by taking some key results in statistical machine learning (including the bias/variance tradeoff), and applying them to the product and engineering tradeoffs we have to make all the time. I'll use examples from a few key adaptive products I admire, and discuss how we're applying these principles to augment conversational commerce at frame.ai.