In the late 2000s Nokia sold more phones than any company on earth, and in some of its meetings criticizing a colleague was off limits. In meetings in front of the chief executive, one senior software manager later told researchers, “there was no way you would accuse someone of not telling the truth.”
Another software manager suggested to a colleague that they criticize a decision by one of the top managers. The colleague said he didn’t have the courage. He had a family and small children.
A third manager remembered how a chart could change on its way to the top: the supervisor would tell you to move the data points to the right, to give a better impression, and then present it to the higher-level executives. “Top management was directly lied to,” the manager said.
Two researchers, Timo Vuori of Aalto University and Quy Huy of INSEAD, collected these stories in 76 interviews. They wanted to know how the world’s biggest phone maker had lost the smartphone race so fast. Executives, they noted, usually blamed weaker technology or complacency. Their answer was fear.
Quiet is not the same as agreed
A team that never argues may not agree. It may just be afraid to.
At Nokia the fear ran in two directions. Top managers feared that admitting their Symbian software was behind would look defeatist to investors, so they kept backing it in public. Middle managers feared the people at the top. Interviewees described the longtime chairman as “extremely temperamental” and several other top managers as highly aggressive. So the honest argument about Symbian’s limits was hard to have anywhere. “I should’ve been much, much more courageous,” one middle manager told the researchers. “Nobody wanted to rock the boat.” Vuori and Huy don’t claim fear was the only reason Nokia fell. In 2013 it agreed to sell its phone business to Microsoft.
Patrick Lencioni named this pattern in The Five Dysfunctions of a Team, published in 2002. The book is a fable, a made-up story about a made-up company, not a report of research. In the wording of his firm, the Table Group, the second of the five dysfunctions is when “the desire to preserve artificial harmony stifles productive ideological conflict within the team.” The word doing the work is “ideological”: the conflict he wants is about ideas, not about people.
Which arguments help
The research is less tidy than the fable. In 2003, Carsten De Dreu and Laurie Weingart pooled earlier studies of two kinds of conflict: about the task, and about each other. Personal conflict went with worse team performance, as expected. But so did conflict about the task, contrary to what textbooks had suggested. Task conflict was less harmful when the two kinds were only weakly correlated.
A larger meta-analysis in 2012, by Frank de Wit, Lindred Greer and Karen Jehn, covered 116 studies and 8,880 groups. They did not find a strong negative link for task conflict. The link was more positive in studies of top management teams, when performance was measured by financial results or decision quality, and where the two kinds of conflict were weakly related.
So the evidence doesn’t say more fighting brings better results. It suggests that arguing about the work can help in some settings, such as top teams making decisions, and more so when it doesn’t come bundled with personal friction. Lencioni’s model itself has had few independent tests. A small 2026 preprint surveying 40 student teams did not find a significant link between their fear-of-conflict scores and grades.
The meeting that ends early
Imagine you lead a planning meeting. You propose a launch date. Four people nod, nobody objects, and you finish ten minutes early.
The easy reading is that the date is fine. Another move is to have everyone write down, before anyone speaks, the one thing most likely to make the date slip, then read the notes aloud. If two name the same supplier, you’ve learned what the nodding hid. If the notes are blank too, the problem may be the room, and the next person to raise a doubt will watch how you take it.
Agents that give in
Nobody at Nokia was a language model. But a team of models can produce the same calm, for other reasons.
In 2025, Andrea Wynn, Harsh Satija and Gillian Hadfield tested multi-agent debate, a setup earlier work had promoted for better AI reasoning. Groups of three models, mixing GPT-4o-mini, Llama-3.1-8B and Mistral-7B, each answered a question, read each other’s reasoning, and answered again for two rounds; a majority vote decided. They used 100 questions from each of three benchmarks and five runs. Debate helped some groups on some tasks, such as math word problems. But on a common-sense quiz, every group scored lower after debating than a vote on their first answers. Among answers that changed, more went from right to wrong than from wrong to right, and a model whose answer no one else shared was the likeliest to drop a correct one. Telling the models they would earn a payoff only for correct answers did not appear to reduce those switches.
These models have no families and no bosses; the authors point to tendencies such as sycophancy and social conformity as possible contributors. What carries over from Nokia is the shape of the failure: the disagreement the group needed was the one least likely to survive the discussion.
You can check your own setup. When several models or agents review a piece of work, save each one’s answer before it sees the others. On questions where you know the right answer, count how often an answer moves from right to wrong after discussion, over several runs, since one run proves little. If the moves keep going the wrong way, take the vote on the first answers.
In the transcripts, that is the moment to look for: an agent with the right answer reads two wrong ones, and changes its own.