What Should an AI Regression Test Protect?

When an AI workflow changes, a test that checks only whether the output still matches an old example can miss the important question: does the workflow still behave safely and usefully when the input is unusual?

For a change to a prompt, model, or tool handoff, which few behaviors deserve a permanent regression test—and who should decide when a changed result is acceptable? I’m interested in how teams balance repeatable checks with room for a genuinely better response.