Evolution of the Databricks Commercial Motion: Building Scalable Intelligence on Databricks

Jason Huang, Stephen Moss17 Jun 2026

Problem

Databricks is a B2B company with more than $5.4B run-rate revenue and more than 10K deals. Finance, sales and strategy gave the CEO three different "revenue booked last quarter" numbers from one Salesforce table. Every team claimed credit for growth, and the claims summed to more than the company's actual growth. Discount approval used a rubric of only deal size and term. The results: metric arguments, over-investment in launches that cannibalized older products, and slow deal approval.

Example from the article: illustrative claims of $600M, $1.0B, $800M and $1.5B from four initiatives against $1.3B of company-wide growth, a 3x over-claim.

Solution in brief

Put one governed deals table behind every number. Then model the revenue that would have existed without the launch and credit the launch only with the gap above that line. For deals, find peer deals with a decision tree and grade a new deal against its peers' discounts.

Questions this case answers
  • How to measure the incremental effect of a product launch against a modeled no-launch counterfactual instead of total growth
  • How to separate cannibalized revenue from net-new revenue when a new product replaces an older one
  • How to measure the incremental revenue of a product launch against a counterfactual instead of total growth
  • How to separate cannibalized revenue from net-new revenue when a new product replaces a legacy one

Tools and methods

  • Governed deals_360 table (Unity Catalog) — one record per deal with agreed definitions and lineage.
  • Counterfactual revenue model — splits a new product's revenue into cannibalistic and accretive parts.
  • Placebo and real runs with bootstrap — effect estimate with a margin of error.
  • Decision tree for peer deals — groups deals with similar discounts.
  • Deal scoring (A / B / F) — auto-approve, modify or auto-reject.
  • Genie, Lakeflow, MLflow — self-serve questions, pipeline scheduling, model serving.
Continue reading — it's free

Google or a one-time email code. No password.