The problem

Atlas asked new users for a workspace name, a team size, an industry, a set of integrations and a plan — before showing them anything. Each step lost people, and the steps that lost the most were the ones collecting data nobody downstream used.

What we did

We audited every field against a single question: does anything in the product read this on day one? Three of the five steps failed that test, so we deleted them and moved the remaining two into the product itself, asked in context at the point where the answer changes what the user sees. The first screen became a query editor with a sample dataset already loaded. Account setup happens behind it.

What we would do differently

We shipped the new flow to all traffic at once. A staged rollout would have told us which of the two changes — fewer fields, or the loaded sample — carried the result. We think it was the sample. We cannot prove it.