Snowflake's Journey: Scaling Enterprise AI Agents from Pilot to 6,000 Users

Snowflake has shared a detailed account of its journey in scaling its GTM AI Assistant, transforming it from a pilot project into a fully integrated enterprise solution for approximately 6,000 users across its sales and marketing organization. Launched in mid-2025, the assistant, powered by Snowflake Intelligence, rapidly became a critical tool, answering over 330,000 questions by the end of the year and significantly aiding users in making faster, better-informed decisions. While previous technical deep-dives focused on the 'how-to' of building such agents, this narrative shifts to the crucial non-technical aspects: team structure, phased implementation, scope management, and the strategic evolution from a Minimum Viable Product (MVP) to a trusted enterprise-wide feature.

From Pilot to Enterprise Deployment

The initiative began in late February 2025 with a focused goal: to develop a Retrieval-Augmented Generation (RAG)-based knowledge assistant. The initial aim was to help Go-to-Market (GTM) teams efficiently locate necessary documents scattered across various fragmented tools. At this stage, the project was supported by a single dedicated Data Scientist. However, by May 2025, a strategic decision was made to expand the scope significantly, aiming for a comprehensive GTM AI Assistant that would not only index knowledge content but also integrate structured sales and marketing data. This pivot necessitated a substantial increase in quality and reliability standards, leading to the expansion of the core team to three to four Data Scientists and the addition of a dedicated Product Manager. From the outset, quality and user trust were treated as non-negotiable, designated as P(-1), recognizing that early user experiences are paramount for adoption. This understanding drove the adoption of a phased rollout strategy: starting small, validating quality and workflows, and then gradually expanding the user base and feature set.

Phased Rollout and Scope Management

Snowflake's approach to scaling its AI agent was deliberately sequential, mirroring its phased rollout strategy. A common pitfall observed in the industry is the premature broad deployment of an agent before it consistently meets quality requirements. The company emphasizes that while prototyping AI agents is relatively straightforward, deploying reliable agents to thousands of users presents a far greater challenge. Therefore, providing teams with adequate time and space to earn early trust is seen as a crucial investment for long-term success. Enterprise environments are inherently complex, with Snowflake's sales and marketing organization alone supporting over 15 distinct personas, each with unique workflows and needs. To avoid the "connect everything and hope users figure it out" trap, the team consciously began with a narrowly defined MVP. This MVP targeted personas where the highest value could be delivered most rapidly. Based on audience size and development effort per persona, the General Availability (GA) launch focused on Account Executives (AEs), Solution Engineers (SEs), and Sales Development Representatives (SDRs). These three groups represented approximately 50% of the target 6,000 users, allowing for maximized early impact while maintaining a manageable scope. Post-GA, expansion was strategic, with additional personas and features being rolled out incrementally without compromising the reliability for the core user group. The agent's capabilities evolved significantly, starting with a knowledge assistant integrated with core Salesforce, product usage, and financial data. Subsequent phases introduced new data sources like marketing data, call logs, emails, partnership data, and web search capabilities, expanding the underlying data layer from 48 tables to 64 tables and increasing columns from approximately 1,400 to over 1,750.

User Activation and Change Management

The successful deployment of enterprise solutions, particularly AI-driven ones, requires significant time, often a full quarter or more. User adoption naturally follows Everett Rogers' theory of the diffusion of innovations, with early engagement typically coming from innovators and early adopters. However, as adoption scales to thousands of users, the reality shifts, with late adopters entering the picture. This can lead to increased expectations and potential frustration if user engagement doesn't accelerate quickly enough. Many AI initiatives falter not due to technological shortcomings, but because users never fully adopt the tools. Ensuring users actually try the AI assistant is paramount, as adopting it into daily workflows represents a significant habit change. Solutions must offer a dramatic improvement—ten times faster or easier—or enable entirely new capabilities to drive adoption, especially among late adopters. The strategy to ensure users not only try but consistently return to the AI assistant involved a deliberate, sequential focus on quality and retention. During the pilot phase, the sole focus was on validating quality and trust, ensuring correctness, reliability, and fundamental utility. In the beta phase, with quality established, the focus shifted to feature completeness and user retention, measuring whether the agent solved enough real problems to warrant weekly engagement. Metrics like a >92% NPS among beta users and a >70% retention rate for weekly active users (WAU) provided confidence for scaling. For the GA phase, the primary question became user awareness and actual usage. This was addressed through close collaboration with the sales enablement team, establishing a dedicated internal product page, clear documentation, user testimonials, and a dedicated feedback channel. Building momentum involved launch emails, live demos, regular mentions by leadership, and weekly adoption reports shared with sales leadership to encourage team participation.

Post-Launch Product Thinking

The successful scaling of enterprise AI agents extends beyond the initial launch; it necessitates a fundamental shift in how data teams operate, moving from project-based work to a product-centric mindset. Unlike traditional data projects that often involve limited post-launch evolution, AI agents are dynamic entities. They continuously interact with users, adapt to new data, evolve with changing workflows, and must keep pace with rapid platform and model innovations. This requires a sustained focus on product thinking, system design, and ongoing team collaboration. Post-GA, Snowflake observed an increase in feature requests, heightened expectations for responsiveness, and the need to balance rapid iteration with maintaining user trust. Bugs required swift resolution, regressions needed prevention, and improvements had to be delivered without destabilizing the user experience. To manage this, the team invested in an agile, sprint-based development process, expanded the team with analytics and backend engineering capabilities, implemented automated testing and CI/CD pipelines for faster, quality-assured deployments, and dedicated resources for platform integrity and development. This approach ensures that AI agents are treated as living products, requiring ongoing ownership, iteration, operational rigor, and continuous investment for long-term success.

Quantifiable Impact and ROI

The impact of the GTM AI Assistant has been substantial. By the end of 2025, it was handling over 35,000 questions per week for more than 2,500 weekly active users, a significant increase from its initial rollout. Usage intensity also rose, with the average number of questions per active user climbing from 8.5 to 14, indicating deeper trust and integration into daily work. Productivity gains were significant, with a conservative estimate of 5 minutes saved per question translating to the annual productivity equivalent of over 65 full-time employees in a 6,000-person company. A detailed analysis revealed that the cost per active user was comparable to standard enterprise productivity tools. Measured against the productivity gains, the assistant delivered an ROI of over 5x, even before dedicated cost optimization efforts. This figure excludes secondary benefits such as reduced analyst load, faster business cycles, and improved decision quality, underscoring the compelling economic case for the solution.

Conclusion

The successful scaling of enterprise AI agents like Snowflake's GTM AI Assistant demonstrates that AI is fundamentally changing how businesses operate, particularly Go-to-Market teams. The era of isolated demos and small experiments is giving way to AI delivering real, measurable impact at scale. Achieving this requires more than just advanced models or sophisticated prompts; it demands the right mindset, disciplined execution, and a strong emphasis on change management. Key takeaways include treating quality and trust as non-negotiable, adopting a deliberate phased rollout, investing heavily in user activation and adoption, and evolving data teams towards long-term product ownership. Companies that embrace these principles are poised not only to enhance productivity but to fundamentally transform their teams' ways of working, securing a significant competitive advantage in the evolving landscape of Go-to-Market execution.