I rehearse with AI before meetings I am nervous about. Pitches, hard conversations, board drills. The first time I did it, I walked in flatter than if I had just prepared alone. The second time, before a different meeting two weeks later, I walked in sharper than I had been in a year.
I almost quit the practice between those two sessions. The rule I use now is simple. Set the model to argue against me, by voice, and run the drill twice before the live session. The first run is a tax. The second one is where the gain shows up.
The CFO who shut it down after one try
A CFO at a mid-market SaaS company in Munich tried this before a budget defense in September. He used voice mode, set to disagree with every line of his plan. Fourteen minutes the night before. He walked into the board less confident than usual. He hesitated on a question he normally owns. He told me after it was the worst session he had run in a year. He blamed the rehearsal and shelved it.
I asked him to run it once more before his next big meeting. He pushed back. I pushed harder. Three weeks later, same setup, different conversation. He texted me that night: "I knew where every one of their punches was going to land."
He almost missed that effect by one drill.
What Columbia measured
Dan Wang and seven colleagues at Columbia Business School ran this exact experiment on 759 MBA students last fall. Preregistered field trial, four instructors blind to assignment. The AI, called CAiSEY, automatically took the opposing position on each case. Students rehearsed by voice, alternating turns, averaging eleven minutes a session.
The headline is what a sceptic would want. After two uses, voluntary contributions rose about 31% of the 0.73-per-session baseline. Comfort speaking up moved. Perceived learning moved. Focus and motivation did not.
The part that matters is the timing.
In the next session, participation was lower after qualifying use at the first assigned opportunity.
β₁ = -0.151, p < 0.05. The first use hurt the next class. The gain only appeared after the second use, and then it persisted across all future sessions.
Most organisations killing their AI-rehearsal pilots are killing them inside this window. One try, bad result, drop it. The architecture of the benefit gets read as a failure of the tool.
Build the practice to argue against the person using it. Make them run it twice before anything they care about. The second drill is the one that pays. The first one is the price of admission.