I used to defend the pre-read ritual. Distribute the deck 24 hours ahead. Ask each attendee to write three bullets on their view. Walk in with the room prepared.
I'm not sure it does anything.
The ritual that moves nothing
A chief strategy officer at a Munich industrial equipment maker, this autumn, prepping the board for a hard call on divesting a legacy division. The room was genuinely split. He had done what he always does. The 40-page deck went out Thursday night. He asked for a three-bullet pre-reflection by Sunday.
The pre-reads came in the morning of the meeting. Half the board had defended positions they wrote at 11 pm. The other half skimmed the deck during the opening. The CSO told me he would try something sharper next quarter. A 15-minute written reflection prompt before the deck.
I kept my face neutral. On the paper I had just read that week, the written reflection prompt would do almost nothing.
What the trial actually measured
Haiwen Li and eleven colleagues at MIT and Google DeepMind ran a randomized experiment on exactly this question. They recruited 1,173 U.S. participants in 391 three-person groups debating contested federal policies. Three arms. A 7-minute dialogue with a non-directive AI partner. A 7-minute written reflection using the same questions. A 7-minute unrelated video.
The group-level Discourse Quality Index rose by Cohen's d = 0.31 under the AI dialogue versus the control video. In rooms with real disagreement, the gap grew to d = 0.38. The written reflection with the identical questions landed d = 0.08, p = .569. Statistically indistinguishable from the YouTube clip.
The 7-minute AI dialogue raised group deliberation quality by d = 0.31. The same reflection questions on paper landed d = 0.08.
What surprised me more. At T3, after the deliberation ended, final positions had not moved any further under the AI condition. All pairwise p ≥ .43 on total opinion change. What shifted was the quality of the reasoning inside the room, not the distance anyone's position travelled.
The caveats are real. The chatbot used gemma-4-31b-it, prompt-engineered as a strictly neutral partner with 2-3 sentence responses and no search. The authors call it a proof-of-concept and say it does not represent off-the-shelf LLMs. Rooms with no disagreement got no measurable lift.
For executives who run two-sided rooms, the operational move is specific. The 7-minute private AI conversation before a hard meeting. Non-directive. Short responses. No browsing. More reflection questions on paper will do nothing. The dialogue is what moves the room. The pre-read never did.