25 years ago I learned that you don't hand an actor a script the morning of opening night and call it a rehearsal. You block the scene, run it quietly, fix what breaks, run it again — and only then does it go in front of an audience.

Most companies are rolling out AI like it's opening night with no rehearsal. Someone in leadership announces "we're using AI now," drops a tool in Slack, and wonders six weeks later why nobody's using it and nothing's actually faster.

Here's what I've seen work instead, from bringing robots and objects "to life" on stage for two decades and now applying the same discipline to AI pilots:

1. Pick ONE process, not five. If you're doing knowledge work — support tickets, onboarding, proposal writing, internal Q&A — there are usually 3-5 places AI could help. Map them all, then commit to exactly one. Trying to fix everything at once is how pilots die.

1. Define the data and the guardrails before you touch a model. What can it see, what can't it see, who signs off before it acts. This is the boring part everyone skips and the part that actually determines whether people trust it.

1. Rehearse it with the people who'll actually use it — not just IT, not just leadership. If the humans doing the work don't help shape it, they'll route around it the day it ships.

1. Ship something small enough to be owned, not something big enough to be announced. A pilot people helped build and can point to as "ours" survives. A platform that got unveiled in an all-hands usually doesn't.

None of this requires an internal AI team or a 6-month roadmap. It requires someone willing to treat the first version like a rehearsal instead of a premiere.

— Ulrike