Three worked examples, one per role

Abstractions are easy to nod along to and hard to build, so here is the whole approach made concrete for a single product, a B2B analytics tool, across its three real roles. The welcome survey asks one question, "what do you want to do first?", with three tappable answers.

The admin taps "set up the workspace for my team". Their activation event is written as a sentence: the admin has invited at least one teammate and connected one data source. The deterministic branch drops them onto an empty state that is an invite panel plus a single "connect your warehouse" card, nothing else. AI layers on top by noticing the admin lingering on the permissions screen and surfacing a short role-permissions explainer unprompted. The personalised welcome email lists three things an admin typically does next. Time to value is measured from sign-up to the moment both the invite and the connection fire.

The analyst taps "explore my numbers". Their event is: the analyst has built one chart from their own data. The branch skips workspace setup entirely and opens a chart scaffold with a "connect a source" prompt where the data would go. If the analyst connects a source but then stalls before building a chart, AI nudges with a one-click "chart this table for me" suggestion built from the columns it can see. The operator taps "get a task done". Their event is one real task completed end to end, so the branch opens the single most common task type, pre-filled with everything inferable, one field from done, and the welcome copy speaks in task language, not metric language. Three roles, three doors, three different first screens, three separately instrumented events, one survey question. That is the entire pattern, and it is exactly the kind of branch-by-intent logic that also powers an AI website chat that qualifies and routes before a person ever signs up.