AI can diagnose a slow PostgreSQL query well, but a human should sign off before touching production.

Query optimization follows repeatable patterns — index analysis, JOIN order, cardinality estimation — but each schema and workload is unique, requiring fresh judgment about trade-offs like index maintenance overhead vs. read speed. The structure is similar across instances, but the specifics demand case-by-case reasoning.

The goal (reduce 45s query to something acceptable) is clear, but 'expected latency improvement' is inherently speculative without running the optimized query against real data. Success criteria are partially crisp (before/after plans) but partially subjective (what improvement is 'enough').

The agent needs live database access to run EXPLAIN ANALYZE, inspect pg_stats, view table definitions, and test materialized view creation — none of which are guaranteed. Without direct DB access or at minimum a full schema dump and query plan output, the agent is guessing rather than diagnosing.

Creating wrong indexes wastes storage and degrades write performance; poorly constructed materialized views can serve stale or incorrect data; and any DDL executed on a 2M+ row production database carries real risk of locking, downtime, or data integrity issues. Mistakes here are not trivially reversible.

A DBA brings contextual knowledge about query frequency, write/read ratios, maintenance windows, and business criticality that an agent lacks. The analysis phase is largely automatable, but deployment decisions and trade-off prioritization genuinely benefit from human expertise.