Summary
Ishita Daga, a machine learning engineer at Tesla building enterprise agents, argues that when a data agent gives a wrong answer, the instinct to upgrade the model or add more context misses the root cause. Her talk identifies three structural problems that actually explain most enterprise agent failures: ambiguity (the agent doesn’t know which of several knowledge bases is authoritative for a given question), staleness (context rots faster than teams update their documentation), and preference (different teams use different metric definitions and query filters, even for the same canonical question).
For ambiguity, Daga proposes a hierarchy of sources of truth: semantic layers (curated KPI definitions and metric calculations) at the top, canonical parametric queries in the middle, and database graphs at the bottom. She recommends starting with the first two, which she estimates address 80% of enterprise queries, before investing in the complex and hard-to-maintain database graph layer.
For staleness, the solution is a context lifecycle with two components: embedding live data sources that are continuously updated (GitHub, CRM tools, dbt semantic layers, Tableau) and a structured feedback loop that logs every correction event — updated metric definitions, new filters, changed processes — and uses those events to trigger automated evaluation runs. Without this loop, she argues, teams have no visibility into whether agent performance is improving or regressing over time. The preference problem, she acknowledges, remains genuinely open and is an active area of research across the industry. A practical framework for any enterprise team deploying AI over internal data.
📺 Source: AI Engineer · Published July 20, 2026
🏷️ Format: Deep Dive
