The first prototype surfaced real friction: the three pieces of information managers actually needed for approval weren't clear, so we added trip dates directly to the notification panel. Managers told us they were too busy to search for alternatives themselves, so version two let them delegate that search to the traveler instead. We also learned people tried clicking Co-Pilot's suggestion card rather than typing, fixed in version two, and that while users tried the microphone icon, several admitted afterward they wouldn't discuss sensitive topics or colleague names by voice in an open office. We also corrected our forecasting model to use only the prior year's data, since reorganizations made five-year comparisons unreliable.

Outcome: a validated, development-ready concept and insight set, built through real iteration with the people who'd actually use it, not assumption.