Case Study: Databricks Lakehouse Pilot

A proof of concept that put scattered CRM, web and order data into one governed lakehouse, and immediately overturned an assumption the business had been running on.

The situation

A manufacturer with data in three disconnected places: a CRM holding deals and contacts, a website capturing enquiries, and an order system holding what was actually sold.

Nobody could answer basic questions that crossed those boundaries. Where do the deals that close actually come from? How much repeat business is there? Which enquiries are real and which are bots? Each system had part of the answer and none of them agreed on what a customer was.

What was done

Built a medallion lakehouse as a pilot. Sixteen bronze tables holding the raw extracts untouched, four silver tables applying the conformance and entity rules, and six gold tables shaped for the questions the business actually argues about.
Resolved the entity question first. Buyers were identified by a salted hash rather than a name, so repeat purchasing could be measured without spreading personal data through every downstream table.
Added a natural-language query layer and a dashboard. So the answers were available to people who do not write SQL, and a daily job rebuilt the silver and gold layers on a schedule.
Kept the credentialed refresh outside the workspace. The token that reaches the CRM stays out of the shared environment, which was a deliberate boundary decision rather than an oversight.

Results

16 / 4 / 6 bronze, silver and gold tables
99% of closed deals came from direct contact, not the web form
60% of web submissions were spam
91% of revenue came from repeat buyers

This is a proof of concept built on a community tier of the platform, and it is described as one deliberately. Its value was analytical rather than operational: three findings that were not visible while the data sat in separate systems. Of 235 closed deals, 232 originated from direct contact and only three from a web enquiry, all of which were lost. Of 780 web submissions, 465 were spam. And 46.8 percent of buyers were repeat customers, accounting for 90.9 percent of revenue.

What transfers to other businesses

The point of consolidating data is rarely the dashboard. It is that a question nobody could ask before gets asked, and the answer changes the plan. Here it meant the funnel was opportunity-led rather than lead-led, and effort aimed at the web form was aimed at three percent of the outcome.

Related services

Client names are not published on this site. Sectors are described instead, and every figure above is taken from the connected analytics property or from a count of work delivered.

Talk about your situation

The first conversation is short and mostly questions. Get in touch and tell me what you are trying to fix.