Reduce Insight Lag by 99% with AI-Powered Voice of Customer Analytics

The Challenge

The company's product and merchandising teams received hundreds of thousands of customer reviews through multiple third-party platforms, including Bazaarvoice, Gladly, and Sierra, but had no systematic process for extracting actionable intelligence from that volume of data. Understanding how specific products performed for targeted use cases, such as high-intensity training or extended daily wear, required hours of manual reading per query. There was no consistent methodology and no mechanism for identifying trends across the full product catalog at scale.

The gap between data volume and usable insight compounded two related problems:

• Sentiment classification was performed on a sampled, ad hoc basis, meaning conclusions rested on anecdotal evidence rather than the complete review corpus.
• Individual analysts could process between 50 and 100 reviews per day, leaving the vast majority of available customer signal unread and unactioned.

The company required a solution that could address three core needs:

• Scalable, automated ingestion: Process the complete review corpus continuously, not in periodic manual batches
• Natural language querying: Enable merchandising teams to ask product questions in plain language without technical mediation
• Consistent, full-coverage classification: Apply uniform sentiment analysis across every review, not a sample

Our Solution

The company partnered with GSPANN to design and build a full-stack Voice of Customer intelligence platform, transforming raw review data into structured, queryable merchandising signals. The strategic decision to build on a modern generative AI stack reflected a commitment to both historical depth and continuous live intelligence, keeping the analyst experience as simple as a natural language conversation.

Review Ingestion and Data Architecture: The team implemented a daily automated ingestion pipeline that pulls customer review data from Bazaarvoice, Gladly, and Sierra into a structured data warehouse built on Amazon Redshift. Each review is processed at ingest time, classified by sentiment label (positive, negative, or neutral), and stored across two purpose-built collections in a Weaviate vector database: one at the individual review level and one at the aggregated product summary level. The pipeline operates through a mature CI/CD workflow, ensuring continuous data freshness without manual intervention.

Natural Language Query and AI Reasoning: The platform enables product and merchandising teams to interrogate the full review corpus using plain English questions, such as which products perform best for high-intensity training or where customers are raising concerns about fabric quality. As an example, for the many users asking questions about product quality, our app answers the questions using information obtained from actual reviews in addition to order details.

Queries are orchestrated through a LangGraph workflow that routes requests to Google Gemini models for interpretation, retrieves semantically relevant reviews from Weaviate, and returns structured, confidence-scored responses grounded in actual review content. Response accuracy is maintained through deterministic model configuration and citation-based grounding, ensuring every answer is traceable to its source reviews.

Our solution used RAGAS metrics during development and achieved ~98% accuracy. In production, responses are grounded using factual signals like order volume, return rates, ratings, vibe analysis, and top reviews. We also display retrieved sources review/product summaries with relevancy scores to ensure transparent and trustworthy answers.

SKU-Level Granularity and Comparative Analysis: The solution operates at the product variant level, supporting analysis by individual SKU in addition to category and sub-category. This granularity enables the merchandising team to identify performance differences within a product line, not just across it, and to compare variants directly on order, return, ratings, and summary, among other data.

Business Impact

• Time to Insight: Reduced from hours of manual review reading per query to seconds via natural language prompt, an approximately 99% reduction
• Analyst Throughput: Scaled from 50 to 100 reviews per analyst per day to processing 200,000 historical records continuously, a 2,000x increase in effective coverage
• Sentiment Coverage: Expanded from sampled and inconsistent classification to 100% automated coverage of all reviews
• Decision Quality: Shifted merchandising decisions from gut-feel and anecdote-driven conclusions to data-backed rankings with quantified confidence scores

Related Services

• Retail Services
• Data & Analytics
• Digital Commerce