Cutting onboarding time by 60% was not an AI problem

The brief was to reduce drop-off during onboarding. The obvious read was that onboarding was too long, so we should cut steps. We measured first, and the data said something less convenient: people were not quitting because the flow was long. They were quitting at the exact moments where they had to go and find an answer somewhere else.

Every one of those moments was a question the product could have answered but did not. What does this field mean. Does this apply to me. What happens if I skip it. Each one sent the user out of the flow, into a help centre or a Slack message to a colleague, and a meaningful share of them never came back.

Adding AI-driven knowledge search inside the flow cut completion time by 60–70% and lifted retention by roughly a quarter. But it would be lazy to file that under "AI works". What actually happened is that we removed the exit. The AI was a delivery mechanism for answers that already existed in our documentation — it just put them where the question was being asked.

That reframing matters for what you build next. If you believe AI fixed onboarding, your roadmap fills up with more AI. If you believe the exits were the problem, you start auditing every flow in the product for the moments where users have to leave to keep going — and you find that most of them do not need a model at all. Some need a tooltip.

The general lesson I keep relearning: instrument the exact step where people leave, not the funnel in aggregate. Aggregate numbers tell you that something is wrong. Step-level numbers tell you what to build. The second one is the only one you can act on before Friday.