AI evaluations must measure customer value, not just performance

Teresa Torres argues that current AI evaluation frameworks focus too heavily on technical performance metrics, such as conciseness or helpfulness, which are insufficient for determining product success. She draws a parallel to the early days of product analytics, where teams initially measured only uptime and speed before realizing they needed to measure customer behavior and value creation. Torres contends that generic, off-the-shelf evaluations fail because definitions of 'good' are context-dependent. For example, a student learning astrophysics requires a detailed, teaching answer, while a professor requires a concise one.

Torres advocates for a cross-functional approach where product teams define specific activation and retention metrics that reflect both customer value and business value. She argues that teams should not outsource the definition of success to generic AI benchmarks but must tailor their evaluation suites to the specific value creation moments of their business. This is similar to how engagement metrics evolved from Monthly Active Users to specific high-value actions, such as sending a document in DocuSign.

The core argument is that AI product success should be treated as an extension of traditional outcome-based product management rather than a new technical discipline. When defining what 'good' looks like, Torres states that it is when the customer did something that created value for the customer. This approach requires moving beyond technical benchmarks to measure the actual impact of AI features on user behavior and business outcomes.