Snowflake Summit: Public Sector Leaders Say Data Foundation Comes Before AI Models

The sessions at Snowflake Summit brought together organizations with very different missions — a university, a defense and government services company, a mission systems contractor and OpenAI's government arm — and the common thread was that each had to fix its data and governance before AI could produce results. The University of Florida's experience illustrates why: when only 10-15% of its data was semantically defined, any insight the institution generated from that data was incomplete and therefore unreliable, so the team built a semantic layer with NaviGator AI and Snowflake components, work it estimates would have taken a human 83 weeks.

Leidos' Alan Sim described a different but related discipline: rather than asking teams what tools they want, he asks what they are trying to solve, and sometimes finds the answer is automation or process re-engineering rather than an agent. Joe Larson of OpenAI for Government framed the same idea from the model side, saying it is more compelling for government to marry the frontier model with its data infrastructure than to simply make models available.

Stephen Moon, Snowflake's public sector field CTO, tied the discussion to mission, and General Dynamics Mission Systems showed what that looks like in practice: its sales quote processing system was ingesting thousands of vendor PDFs with varying formats, layouts and terminology, a training-based approach failed, and zero-shot prompting with Snowflake Cortex AI took it from 50% accuracy to 85% accuracy in its recent go-live to production. The outcome, the article says, was measured in speed, accuracy and cost savings on a workflow that directly affects mission delivery. Whether other agencies can replicate those results appears to hinge on how much of their own data foundation, governance and mission alignment are ready before they scale.