Why AI Index Suggestions Need Experimental Validation Before Production Use

Language models can generate plausible database index recommendations, but they have no visibility into a system's actual I/O patterns or query planner behavior, making blind adoption risky. Accepting an AI-proposed CREATE INDEX statement without testing can introduce write regressions in production, even when the model's explanation sounds technically convincing. The recommended approach is to treat any suggested index as a hypothesis, pairing it with a controlled before-and-after benchmark on an isolated copy of the data. A simple test harness — capturing query plans and wall-clock times before and after applying the candidate index — can reveal whether the change delivers real gains or merely sounds reasonable. This methodology applies across database engines including SQLite, PostgreSQL, and MySQL, with only minor adjustments to the plan-capture command.