Building GTM AI Agents: Lessons from Deploying to 6,000 Users — Sait Izmit, Snowflake
AI Engineer World's Fair 2026 · 20:39
← All organizations
Enterprise data and AI
Snowflake
Snowflake provides a managed cloud platform for enterprises to store, process and share data, build applications, and run AI workloads. Data engineers build ingestion and transformation pipelines; analysts query data with SQL; developers use Snowpark to run Python, Java and Scala code. Its Cortex and Snowflake ML offerings connect AI development to enterprise data: Cortex uses large language models to analyze unstructured content and answer questions, while Snowflake ML supports custom model development, deployment, feature stores and model registries.
Founded in 2012 by Benoit Dageville, Thierry Cruanes and Marcin Zukowski, Snowflake is led by CEO Sridhar Ramaswamy. Its defining database architecture separates persistent storage from independently scalable compute clusters called virtual warehouses. Multiple teams can process the same stored data using separate compute resources. The design also makes semi-structured formats such as JSON accessible through SQL. Snowflake’s 2016 architecture paper received the 2026 SIGMOD Test-of-Time Award.
Snowflake earns product revenue primarily through customers’ consumption of compute, storage and data-transfer resources. In September 2026, the company reported more than 14,500 customers worldwide. Its platform extends beyond internal analytics: organizations can share selected database objects with other Snowflake accounts, distribute data and applications through its Marketplace, and use data clean rooms to permit defined analyses without granting unrestricted access to shared data.
Explore the recordings
This guide covers Snowflake’s supplied recording archive: one talk about an internal go-to-market assistant. The paths below explore different parts of that recording. Scale, results, platform names, and capabilities are claims reported in the talk, not verified facts about Snowflake’s products today.
Start with evaluation and user trust
For teams deciding what an agent should answer, begin with Sait Izmit’s deployment lessons. He describes deriving 150 evaluation questions from the sales process, finding roughly 50% initial accuracy, and narrowing coverage to improve reliability during users’ first interactions. Follow this thread into the accuracy-focused pilot and staged rollout to see how evaluation informed launch decisions.
Follow the rollout into adoption
For internal product owners, use the same recording to examine the path from a pilot to a 600-person beta and then general availability. Izmit reports beta retention above 70% and deployment to roughly 6,000 users. His discussion of demonstrations, leadership sponsorship, and adoption measurement offers a path through the organizational work after launch; those reported results describe this deployment rather than establish general benchmarks.
Explore architecture, feedback, and expanding expectations
For builders operating an agent beyond launch, follow Izmit’s account of system evolution: simple initial instructions and services grew to include evaluations, CI/CD, skills, MCP integrations, memory, scheduling, and new interfaces. Pair his discussion of conversational access giving way to workflow automation and personalization with his recommendations for LLM-assisted analysis of interaction logs. The Q&A describes a platform he calls Snowflake CoWork, formerly Snowflake Intelligence in his account; treat that naming and the described capabilities as recorded statements, not confirmation of current branding or availability.
1 talk
Newest first1 speaker at AIE
Affiliations reflect their AIE appearances, not necessarily current employment.