Track workflow starts, completions, cancellations, retries, corrections, feedback, and meaningful outcomes. Join them to privacy-safe trace and version identifiers.
Avoid logging raw prompts or images by default. Measure latency and cost by feature, then compare them with completion and retention. Analytics should answer whether AI helped the user finish a job—not merely whether the model returned text.