Industry: Commercial Manufacturing

Timeline: 6 Weeks

AI-Powered Product Intelligence & Sales Enablement Platform

Team: Tempered AI, in partnership with a leading technology consulting firm

4-1-edit
March 24, 2026
4-2 edit

Executive Summary

A century-old commercial tile manufacturer with a catalog of over 170,000 products needed a smarter way to work with its own data. Tempered AI joined the project to build three AI-powered solutions on Databricks in roughly six weeks, turning a process that used to take weeks into one that takes minutes, while making product specs and sales data searchable across the entire organization.

About the Client

Our client is a well-established U.S. manufacturer. They serve specialized environments like hospitals, schools, and large commercial buildings, where each project requires a precise combination of tile materials, finishes, dimensions, and performance ratings (think humidity resistance, acoustic ratings, and sanitary compliance).
Their product catalog had grown to over 170,000 SKUs, pulled from multiple internal pipelines and partner systems over the years. The result was a wide, complex dataset with over 150 attribute columns per product and a lot of inconsistency baked in.

The Challenge

The client was sitting on a goldmine of product and sales data but had no practical way to use it. Three problems stood out:

Inaccessible product documentation

Thousands of technical specification PDFs were stored in a database that nobody could efficiently search. If you needed a specific detail about a product, you had to find the right document and read through it manually.

Fragmented sales data

 Historical sales information existed across regions and business units, but there was no single place to query it. Sales teams couldn't easily pull performance data by region, product line, or time period.

A slow recommendation process

When a sales rep met with a new client, they would go back to the office and spend two to three weeks manually researching which tiles and grid systems would be the right fit. Then they'd return with a handful of suggestions. It worked, but it was slow and hard to scale.

On the technical side, the data itself was a challenge. A table with 170,000 rows and 150+ columns, heavy null values across many fields, and column names that were entirely internal and undocumented made the early stages of the project demanding.

The Solution

Before writing a single line of code, the team spent significant time with the client's internal data lead, a person who had been with the company for over 20 years and understood the dataset inside and out. That discovery process shaped everything that followed.
The team ultimately built three distinct products, all running on Databricks.
Product 1: Product Knowledge Chatbot
A conversational AI assistant that let users ask plain-language questions about any product in the catalog. It pulled from cleaned and structured product tables, and also included a RAG (Retrieval-Augmented Generation) layer that indexed the technical PDFs so their content became searchable for the first time. A user could type something like "What are the acoustic specs for this tile?" and get a clear answer in seconds.
Product 2: Sales Intelligence Chatbot
A chat interface that gave sales reps access to regional sales data through natural language. Access was filtered by role and territory automatically, so someone in the northeast could query their region's data without seeing anything they shouldn't. Questions like "How did Product X perform in Q1 last year?" could now be answered on the spot.
Product 3: Product Recommendation Engine
This was the flagship product and the most technically complex piece. A sales rep could describe a client's situation (a small school in Miami, for example) and the system would figure out what that meant: humid climate, bright colors for light reflection, specific safety and acoustic requirements. It would then pull from the spec PDFs, query the structured product tables, and run everything through a similarity model built across the full 170,000-product catalog to surface the top 10 compatible tile and grid combinations.
What previously took two to three weeks of manual research now came back in five to ten minutes.

Technical Highlights

The stack was built on Databricks for all AI and data work, with Azure Functions and an Azure Web App handling the integration layer.
One of the more thoughtful decisions was around delivery. Rather than building a separate application that users would need to learn, the team integrated all three tools directly into Microsoft Teams, which the client already used daily. Authentication was handled through existing Microsoft accounts, with role-based access logic running in the background. Users just opened Teams and started asking questions.

The Result

Sales recommendation time went from two to three weeks down to under ten minutes
Over 170,000 products became searchable through natural language for the first time
Three production-ready AI tools delivered in roughly six weeks
Full adoption from day one since everything lived inside a tool the team already used
Automated security controls with no added complexity for end users
Wrapped up with knowledge transfer sessions and a business review presented to 30+ stakeholders

Why It Works

The team came in with a point of view. Rather than asking the client what they wanted and building exactly that, they listened, did the research, and came back with recommendations. When it became clear mid-project that a standalone web app wasn't the right fit, they shifted to the Teams integration without losing momentum.
Regular demos throughout the build kept feedback flowing and meant there were no surprises at the end. And leaning on the client's internal data expert early on made everything downstream faster and more accurate.
Interested in Self-Service Analytics for Your Organization?
Whether you're in healthcare or another highly regulated industry, we can implement analytics platforms that balance accessibility with governance.
Start a Conversation
4-1-edit (3)
4-1-edit (2)