How Narwal Built a Snowflake AI Agent to Transform Product Reviews into Instant, Governed Insights

Background

Product reviews tell a story, but for a global manufacturer managing 50+ product lines across multiple retail channels, that story was trapped; until an AI agent built natively on Snowflake made it accessible.

Business and product teams were asking real questions: Which product lines are drawing the most negative feedback? What specific issues keep surfacing in reviews? How has sentiment shifted over the past quarter? Getting answers meant looping in technical teams, running separate queries across disconnected tools, and waiting. The insight was there. The path to it was not.

Answering these questions required manual SQL queries across disconnected tools, creating a 3-to-5-day technical bottleneck. Narwal was brought in to build a Snowflake-native AI agent capable of processing natural language queries, automatically routing requests, and delivering instant insights within a secure environment.

Challenge

Building an enterprise-grade GenAI solution inside a governed data platform introduced four primary hurdles:

• Divided Data Paths: Qualitative context lived in unstructured reviews, while quantitative metrics lived in structured tables. No single interface connected both.
• Manual vs. Automatic Routing: Business users needed instant answers without having to manually select retrieval tools or reformulate queries.
• Strict Data Governance: Due to compliance standards, customer review data could not leave the client’s secure Snowflake environment via external APIs.
• Semantic Accuracy: Without a well-defined business data model, the AI agent risked generating inaccurate or misleading metrics.

Solution

Narwal designed a unified, Snowflake-native architecture that automatically interprets user intent and orchestrates the backend response using native LLMs like Llama 3.

• Agentic Orchestration with Semantic Routing: The core of the solution is a Snowflake Agent configured to classify incoming queries by type. Retrieval-oriented questions go to Cortex Search. Analytical questions go to Cortex Analyst. Mixed-intent queries trigger both. The routing logic is embedded in the agent, not bolted on externally.
• Cortex Search for Qualitative Retrieval: Narwal configured Cortex Search over the client’s product review corpus. When users ask open-ended questions about customer sentiment, the agent pulls the most relevant review content and surfaces it as grounded context.
• Cortex Analyst for Structured Reasoning: For metric-driven questions, the agent hands off to Cortex Analyst, which runs against structured product data. This enables quantitative outputs like sentiment scores, trend comparisons, and category-level breakdowns without leaving the Snowflake environment.
• Semantic View as the Analytical Foundation: Narwal defined a Semantic View over the client’s product data to give the agent a clean, business-aligned model for analytical queries. This layer ensures that computed answers reflect how the business actually thinks about its data.
• Single Governed Workflow: The full pipeline, from query intake through routing, retrieval or analysis, and response, runs inside Snowflake. No data leaves the platform. No external APIs carry sensitive review content.

Outcomes

• Core Agent Architecture Delivered: The agentic orchestration model is built and validated. Semantic routing across Cortex Search and Cortex Analyst works as designed, with query classification happening automatically based on user intent, positioning the client within the 50% of business decisions AI agents are expected to augment or automate industry-wide by 2027.
• Qualitative and Quantitative Sentiment in One Place: For the first time, the client can ask a single question and receive an answer that draws on both retrieved review context and computed data, without switching tools or waiting on a technical team.
• Data Stays in Snowflake: The governed architecture ensures the client’s review data never leaves the platform. Access controls, data residency, and auditability are maintained end to end.
• Validated Prototype, Production-Ready Foundation: The advanced prototype has been tested across a range of query types and edge cases. The architecture is stable and designed for scale, giving the client a clear path from POC to full deployment, built on a Semantic View that reflects the same semantic rigor shown to lift GenAI model accuracy by up to 80%.
• Extensible Beyond Sentiment: The semantic routing pattern Narwal built is not limited to product reviews. The same agent framework can be extended to other Snowflake data domains, making this engagement the starting point for a broader agentic AI capability inside the client’s environment.

Conclusion

Product intelligence should not require a technical translator. Narwal built the layer that removes that dependency, an agent that lives inside Snowflake, understands what a question is asking, and finds the answer through the right path, whether that means surfacing a relevant review or computing a trend.

For this manufacturer, the result is a governed, intelligent interface over data they already own. The questions that used to take days to answer are now a query away.

Ready to Scale Your Enterprise GenAI Strategy?

Don’t let valuable business insights remain siloed. Partner with Narwal to turn your Snowflake environment into an intelligent, agent-driven layer for enterprise decision-making.