Summary
Sait Izmit, responsible for internal AI tools at Snowflake’s sales organization, presents hard-won lessons from deploying a GTM AI assistant to roughly 6,000 users — a system built on Snowflake Cortex that has now handled over one million questions and answers around 40,000 per week. The talk is unusually candid about the gap between launch excitement and sustainable adoption.
Izmit describes starting with 150 real seller questions before writing a line of agent instructions, running the first test, and hitting 50% accuracy. That disciplined eval-first approach — his team’s internal motto is “quality is P minus one” — became the foundation for iterative improvement. He walks through the activation curve: getting sellers to try the product for the first time is the hardest problem, and each bad answer in the first five interactions is extremely costly to recover from.
The talk also addresses what he calls the “collapsing of the wow factor” — the inevitable moment when a transformative tool becomes habitual, and users start demanding more. His framing of the maturity arc resonates: teams move from “talk to your data” (democratizing access from dashboards and analysts) to “automate my workflows” via MCP integrations, to building team-specific skills and custom agents. Izmit’s perspective as Snowflake’s own customer zero for Cortex products gives the session an insider quality that goes beyond generic enterprise AI advice.
📺 Source: AI Engineer · Published August 26, 2026
🏷️ Format: Workflow Case Study
