How Furniture.com Used Databricks Mosaic AI to Deploy 8 Production AI Models and Double Conversion Rates

Furniture.com connects shoppers to furniture from more than 70 retail partners through a single cart experience, making product discovery and decision support central to its value proposition. When shoppers are comparing dozens of similar sofas or dining sets, the quality of AI-driven product information — summaries, collections, recommendations — directly determines whether they convert.

The challenge was scale: processing millions of product reviews and enriching hundreds of thousands of product listings required a data infrastructure that could support rapid ML experimentation and reliable production deployment. A five-person data and AI team needed to iterate quickly and ship models to production without the overhead of managing infrastructure.

Furniture.com standardized on Databricks Data Intelligence Platform, using Delta Lake as the foundational storage layer, Unity Catalog for data governance and model versioning, MLflow for experiment tracking, and PyFunc models for flexible inference. The team built AI review summary generation using Mosaic AI LLMs — processing 4.5 million product reviews to generate summaries that clearly communicate product features, materials, and customer experiences in natural language. They also developed NLP-powered collections that group products by semantic meaning rather than rigid category taxonomies.

The results were measurable at the shopper level. AI-generated review summaries engage 95% of shoppers who encounter them. NLP-powered collections drive a 26% conversion boost. The single most impactful AI intervention — AI review summaries — delivered 2x conversion compared to pages without them. The entire suite of 8 production AI models was deployed by a team of five within twelve months.

As Yasel Garces, Associate VP of Data and AI at Furniture.com, noted: "To adapt AI to your specific use case and deliver something meaningful to users, you need significant preprocessing." The team's quality-first approach included dedicated human QA review of AI outputs before shipping to production.