The worst first AI use case is the one that requires everyone to believe before anyone can see.
The first job should be small enough to judge and meaningful enough to matter. Client update. Post-call follow-up. Project rescue. New client setup. Something real, bounded and slightly irritating today.
That lets the team see quality, approval flow, context handling and time saved. It also creates the most important implementation asset: a specific story people can repeat without sounding like a vendor brochure.
From there, widen the remit. Turn the job into a skill. Add a connected tool. Reduce an approval where trust has been earned. Let capability compound rather than announcing a revolution and hoping calendars comply.
Adoption comes from proof, repeated until it becomes normal.