Metrics framework for an AI feature: user value, output quality, cost and safety

You are the product manager for a feature or product whose core output comes from an AI model. How would you define the metrics that show whether it is succeeding? ```hint Look past usage An AI feature can be used heavily while its outputs are wrong, unsafe or expensive to produce. Ask which numbers would reveal those failures, and which would only count activity. ``` ### Clarifying Questions - Which product should anchor the answer: an answer engine that writes a sourced response to each query, an assistant built into an existing app, or a tool that generates images or text? - Is the feature newly launched, so trial and repeat use matter most, or established, so retention and conversion to a paid plan matter most? - Can users rate or flag outputs (thumbs up or down, regenerate, report), or must quality be inferred from what they do next? - Should the cost of running the model on every request count toward success, for example on a free tier? ### What a Strong Answer Covers - Metrics derived from the product's goal and lifecycle stage rather than a list of model statistics. - A north-star metric that reflects users getting value from the AI output, and why raw usage is a weak proxy for it. - Output quality measured both offline (correctness, groundedness, hallucination rate on a labeled set) and online (explicit feedback and behavioral signals), with the biases of each. - Guardrails for safety and trust, latency, and inference cost per request, and how they trade against the north star. - How the metric set is used: gating model changes, running experiments, and cutting results by query type, model and user segment. ### Follow-up Questions - After a model update, the thumbs-down rate fell, but so did the share of answers that received any rating. What could explain it, and what would you check? - How would you estimate the hallucination rate at scale when most users never open the cited sources? - Leadership wants a single number to report every week. Would you give them queries per user? If not, what would you give them and why?