Process mapping with AI
Turn interview fragments into a swimlane map with AI drafting and humans verifying — because the map is a set of claims, and unverified claims make fictional processes.
Your investigation found what broke. Fixing it — and preventing the next one — requires seeing the process: the sequence of steps, handoffs, systems, and decisions between 'customer clicks buy' and 'refund issued.' Processes are where causes hide (the fragile flag nobody owns lives in a process gap) and where fixes get installed. The BA's map of choice is the swimlane: rows are actors (customer, support, warehouse, finance, systems), boxes are steps, arrows cross lanes at handoffs — and handoffs are where processes break.
AI drafts the map; interviews correct it
From these interview excerpts and this ticket-flow description [paste], draft Harbor Lane's returns process as a swimlane map in text form: lanes for customer / support agent / warehouse / finance / systems, numbered steps in sequence, handoffs marked with '->', decision points marked with '?'. Where my sources are silent or contradictory about a step, insert [UNKNOWN: what to ask] rather than guessing. List all UNKNOWNs at the end as an interview checklist.
The [UNKNOWN] instruction is the whole trick. Unconstrained, AI produces a beautiful, complete, plausible process map — including steps it invented. A map that marks its own gaps is a discovery tool; a map that hides them is a liability.
- Walk the map with the people in it. Fifteen minutes with a support agent ('is this actually what you do?') surfaces the differences between the official process, the believed process, and the actual process. All three exist everywhere; the gaps between them are your findings.
- Annotate with data, not adjectives. Volume per step, wait time per handoff, error rate per decision — from Module 2's extracts and ticket data. 'The warehouse inspection step is slow' is an opinion; '4.2 days median, 60% of end-to-end time' is a map annotation that survives a meeting.
- Hunt the orphan steps. Every step needs an owner. Your investigation already found one orphan: who maintains the fragile flags? 'The catalog person, years ago' is process-speak for 'nobody' — and unowned steps are where quiet failures compound. Ask AI to list every step whose owner is unclear from the map; it's a mechanical check, and it catches real gaps.
A swimlane in numbered-text form (like the prompt produces) is pasteable into tickets, emails, and AI conversations, versionable, and readable without a license to a diagramming tool. Make the pretty diagram for the readout if the audience expects one; keep the text version as the working artifact. The map's value is its claims, not its rectangles.