The question I keep hearing from executives is whether to invite the agent to the meeting. That is the wrong question.
The right one is how many voices in the room are allowed to be AI. And the answer bends in a shape nobody expects.
The chief of staff who was proud of her huddle
Last month I was on a coaching call with a chief of staff at a mid-market B2B SaaS company in Amsterdam. She had rebuilt the Monday strategy huddle. Four humans in the room. Two AI copilots at the table, one running market intel, one tracking competitors. Six voices total.
She told me the team was converging on decisions faster than ever. She was proud of it. I asked her to read me the last three decisions out loud.
She read them. Then she went quiet.
The decisions were coherent. They were also foreign. Nothing in them sounded like her team. No customer story. No one arguing for a client they had lost sleep over the week before. The room had converged. Nobody in it had actually decided anything.
I did not have the language for what was happening until this week.
What the paper measured
Chen, Liu, Hu and Li at Tsinghua and Northeastern (arXiv:2609.02122, published September 2) ran a between-subjects experiment on a collaborative description game. They put 79 humans and Qwen2.5-VL-32B-Instruct agents together, varying the proportion of AI voices across five conditions from 0% to 75%.
The effect is a U-shape.
A small AI presence at 12.5% strengthened human consensus, +8.0% over the pure-human baseline. A one-third AI presence collapsed it, a 23.1% drop (z=-12.326, p<0.001). A three-quarter AI presence restored consensus. The restored consensus was agent-shaped:
concreteness 2.986 vs 2.724, analogical ratio 0.802 vs 0.050, event-framing ratio 0.385 vs 0.000
The humans, in a room mostly full of AI voices, agreed on abstract ideas stripped of the real-world analogies that make organisational language mean anything. At the 33.3% mark, agents contributed 60.7% of the final vocabulary. And humans, the abstract notes, "initially resist adopting expressions from partners perceived as AI but gradually yield to conformity pressure."
The Amsterdam huddle was sitting exactly on the 33% cliff. Two agents, four humans. The team converged faster because they were being pulled toward a language none of them would have written and none of them were fighting for.
Ask a different question. How much of the vocabulary of this decision am I willing to hand over? A little AI voice sharpens the group. A third breaks it. A majority-AI room agrees on ideas that never lived anywhere.
Watch what your team stops saying. That is the meter.