Easy as a green run: How Vail Resorts built an AI assistant to automate personalized recommendations

Vail's AI assistant represents a critical inflection point for analytics engineering: personalization at scale now requires real-time feature serving and low-latency inference, not batch workflows. This fundamentally changes our infrastructure decisions. We're moving beyond warehouse-centric architectures toward hybrid systems where transformation happens closer to the application layer—think feature stores, streaming pipelines, and edge inference. For data teams, this means adopting tools like Tecton or Feast isn't optional anymore; it's operational necessity. The operational implication is brutal: you can't govern transformation in a single dbt project when recommendations need sub-100ms response times. We need to decouple our analytics layer from our operational ML layer. The broader pattern is clear—every company building customer-facing AI products will face this infrastructure crisis. My recommendation: audit your current recommendation systems now. If you're still running nightly batch jobs feeding into dashboards, you're already behind. Start experimenting with streaming aggregations and feature platforms immediately, even if it means parallel infrastructure initially. The tooling maturity is finally there.