The tempting first pilot is always the exciting one — the customer-facing feature, the flagship use case. It's also the one where a visible failure does the most damage to internal trust in the whole initiative.
A boring first pilot — internal document search, ticket triage, meeting summaries — has a much smaller blast radius when something goes wrong, and it usually does go wrong at least once early on.
Boring pilots also generate the operational muscle a team actually needs: who monitors output quality, what the escalation path looks like, how often the model needs re-checking. That muscle transfers directly to the exciting use case later.
Teams that start with the flagship use case tend to either quietly shelve it after a rough first month, or ship it anyway under pressure — both outcomes are worse than a boring pilot that just works.
"Boring" here doesn't mean low-value — internal search and ticket triage save real hours. It means low-drama: a bad day doesn't reach a customer, and doesn't become the story people tell about why AI "didn't work here."