When I joined the Atlassian Growth team, Atlassian Intelligence had a telemetry problem. Not a product problem. The features were solid. The problem was that nobody really knew who was using what, or why the numbers had plateaued. That's where a technical program manager either earns their keep or doesn't.
The first thing I did wasn't a roadmap review. It was a data audit. I spent two weeks mapping what signals actually existed, what was being tracked versus what was being reported to leadership, and where the gaps were. Adoption numbers that look flat are usually a measurement problem before they're a product problem. In this case we were counting sessions that touched the feature, not sessions where someone got value from it. Very different number.
Once we had accurate instrumentation, the adoption story changed fast. We'd been undercounting by about 30 percent. But even the corrected number needed to grow, so we moved to the real work: getting the feature into workflows people were already in. For Atlassian Intelligence in Jira that meant the Slack integration. Same surface engineers already lived in all day. The integration was real engineering effort, but the program challenge was getting three teams to sequence their work so none of them ended up blocked.
That sequencing problem is where enterprise AI program management actually lives. The model works. The gap is cross-functional coordination at the pace AI cycles demand, which is faster than most legacy program structures were built for. I restructured the cross-team cadence: cut the weekly all-hands in favor of async updates and two tighter working sessions, one on blockers and one on dependencies.
We hit 60 percent-plus adoption on the integrated surface and lifted Confluence document engagement 277 percent through a parallel Slack strategy. The lesson that stuck: AI adoption is a program management problem wearing a technology costume. Alignment, sequencing, and measurement move the number. The model doesn't.