11AI applications often add embeddings and model calls to an existing data path. That makes familiar disciplines—query tuning, indexing, lineage, and workload management—even more important, not less.
Latency is cumulative
An agent may perform several retrievals and tool calls before responding. A slow query repeated across that path can dominate the experience. Teams should measure end-to-end latency and attribute it to each component.
Query plans, appropriate indexes, bounded result sets, and caching can often improve the experience more predictably than changing models.
Business logic needs one home
When metric definitions differ between dashboards, warehouses, and agent tools, the model may surface inconsistent answers. Shared semantic definitions and governed views reduce that ambiguity.
The goal is not to move every calculation into SQL. It is to make authoritative logic discoverable, testable, and reusable.
Design for a mixed workload
Vector search, analytical queries, operational lookups, and ingestion compete for resources differently. Workload isolation and realistic load testing prevent an AI feature from degrading the system it depends on.
Strong AI architecture still begins with strong data engineering.