Why We Rebuilt Our Onboarding Flow
When we shipped the original onboarding wizard two years ago, five steps felt reasonable: connect an account, name your workspace, invite a teammate, pick a template, and confirm. In practice, we watched session recording after session recording of people abandoning at step three.
The team spent a month just watching. Not designing, not building - watching. We ran fifteen moderated sessions and pulled aggregate funnel data across roughly four thousand signups. The pattern was consistent: people wanted to see a working automation before they were asked to invite anyone else.
So we flipped the order. The new flow drops you straight into a pre-filled example automation, lets you run it once, and only asks about teammates after you've seen something work. Inviting a teammate moved from step three to an optional prompt after your first successful run.
Completion rate for reaching a first successful automation went from 61% to 84% in the six weeks after launch. Invite volume, somewhat surprisingly, went up too - people invite teammates to something they already trust.
None of this required new technology. It required sitting with the discomfort of watching real people struggle with something we had built, and being willing to reorder five steps we'd stopped questioning.
The hardest part wasn't the redesign - it was agreeing to stop defending the old flow. Everyone on the team had a theory for why step three was necessary, usually some variation of 'if they invite a teammate early, they're more committed.' The data never supported that theory, we just hadn't looked hard enough to notice.
We also learned to distrust our own dashboards. Our internal funnel metrics showed a healthy 74% completion rate for months, because we were measuring from 'account created' rather than from 'wizard started.' A meaningful chunk of people were never entering the wizard at all, and that gap didn't show up anywhere until we watched real sessions.
Support tickets dropped too, which we hadn't predicted. A recurring category of ticket was some version of 'I don't think my automation is actually running,' which mostly turned out to be people who'd never made it to a first run at all. Fewer people stuck felt indistinguishable, from the support queue's perspective, from fewer people confused.
We're not treating this as finished. The new flow still asks for a workspace name before showing anyone anything, and we suspect that's the next thing worth questioning - but we're going to watch fifty more sessions before we touch it.