What This Session Is About
Every week, thousands of customer satisfaction survey responses flow into Snowflake's support systems. Before this project, a team of people manually reviewed each line — classifying feedback as positive, negative, or neutral, and tagging it by category (product issue, service quality, documentation gap, etc.). The process took weeks and introduced inconsistency: two people could classify the same feedback differently.
Rajanikant built an automated pipeline using Snowflake Cortex AI's classification feature that runs every Sunday night, processes the week's feedback, and delivers structured insights to leadership by Monday morning. He shared the journey — including the pitfalls — in this session.
The 5 Best Practices
Before training any classifier, Rajanikant's team discovered non-English text, blank responses, single-word replies, and noise. They used LLMs to translate non-English feedback, filtered out blanks and meaningless ratings, and removed extremely short responses. Accuracy jumped from 60–70% to 92–93% after cleaning alone.
Rather than rebuilding from scratch, they wrapped Snowflake Cortex AI around existing data pipelines. The AI classify feature needed 20 labeled examples per category — they provided those from past human-reviewed data and let the model generalize. The workflow runs on a schedule without manual triggers.
Even though the pipeline is fully automated, someone on the team spot-checks a random sample weekly, validates classifications, and feeds corrections back as new labeled examples. The model improves continuously. This isn't optional — customer behavior changes, and the model must keep pace.
Customer language and priorities shift. A phrase that indicates a product issue this quarter might indicate service quality next quarter. Rajanikant scheduled periodic re-evaluation with fresh labeled data to prevent drift. Treating a deployed model as "done" is how accuracy silently degrades.
For customer-facing AI (chatbots, recommendation engines), Rajanikant argued that companies must clearly disclose what data is shared with AI models and give users control. Regulated industries (healthcare, finance) carry particular risk — the session referenced a high-profile incident where government workers pasted sensitive data into ChatGPT.
Key Insights
- 01Weeks of work compressed to hours. The nightly pipeline processes an entire week of CSAT responses and delivers classified, categorized results to leadership dashboards before Monday morning standup. What used to take human reviewers 2–3 weeks now runs automatically while everyone sleeps.
- 02Accuracy improved from 60% to 92–93% through data quality, not model size. The Snowflake Cortex AI_CLASSIFY feature is not the most powerful model available — but paired with clean training data and diverse labeled examples, it consistently outperformed the earlier human-reviewed process on consistency.
- 03The closed feedback loop changed how engineering prioritizes. Previously, feedback reached product teams weeks after the fact — often after they'd already moved on. Now, product and engineering receive weekly classified signals. Features get prioritized while customer memory of the issue is still fresh.
- 04Snowflake Cortex AI runs inside the data warehouse, RBAC-protected. A key architectural advantage: the AI classification runs inside Snowflake, where data is already governed by role-based access control. No data leaves the environment to be processed by an external API.
- 05Hyperpersonalization and predictive CX are the next frontier. Rajanikant described proactive support patterns — where AI detects a network issue and dispatches a technician before the customer notices — as the direction the entire industry is moving. Reactive support is becoming a table-stakes floor, not a differentiator.
- 06AI-powered text-to-SQL is bridging the gap between BI and natural language. Snowflake's Analyst feature builds a semantic model over your data and converts natural language questions to SQL, showing the generated query before executing. It removes the dependency on data teams for ad-hoc questions.
Before AI, it used to take like weeks to go through thousands of feedbacks line by line. Now the system automatically classifies everything and generates a request for product and engineering — and they are prioritizing it in their backlog.
From the Q&A
How do you measure success of the workflow?
Three metrics: (1) time spent on classification before vs. after — weeks to hours. (2) Accuracy — went from 60–70% to 92–93% after data cleaning and diverse training examples. (3) Time-to-action for product and engineering — how quickly features requested in feedback make it into a release.
Are you still improving the model?
Yes. The human spot-check loop feeds new examples back each week. The model is retrained on a schedule to reflect how customer language and priorities evolve. It's never truly "done."
What about over-personalization — can you go too far?
This came up around customer-facing AI. There's a fine line between helpful personalization and making customers feel surveilled. Rajanikant's guidance: stay within your product's data (purchase history, support history) and give customers explicit opt-out controls. Anything beyond product data starts eroding trust.
How do you handle change management when introducing AI to teams?
With large organizations, some pushback is inevitable. His approach: frame AI as a co-pilot that reduces toil, not a replacement. Provide hands-on training so teams experience the time savings directly. The teams that were most skeptical became the strongest advocates once they saw the Monday dashboards.