“Given a free choice of members and the need to form a high-powered management team, how should we go about it?”
Meredith Belbin had the means to find out. The managers who came on courses at Henley Management College in England played a business game, the Executive Management Exercise, in which they were divided into companies that competed with one another. Belbin’s researchers watched how each company worked. According to the Belbin company’s account of the research, every participant also completed a battery of psychometric tests, and one of them, the Critical Thinking Appraisal, measured mental ability.
So the researchers did the obvious thing. They built companies that, in Belbin’s phrase, hoarded mental ability, and called each one an Apollo company. The other course members, he wrote, reacted immediately. They recognized the Apollo company for what it was.
“The Apollo team finished last,” Belbin wrote.
Twenty-five Apollos
One result could have been bad luck. “Promising experiments always need to be repeated,” Belbin wrote, and his team repeated it. By his count, of twenty-five companies built to the Apollo design, only three won. The most common finish, out of eight, was sixth.
Watching them, the researchers saw where the time went. Apollo members, by Belbin’s account, had spent much of it “engaged in abortive debate,” each trying to bring the others round to his own well-argued view.
A team may need a mix of ways to contribute more than a room of the cleverest people.
That was the argument of Belbin’s 1981 book, Management Teams. He and his researchers sorted the behaviors that seemed to separate winning companies from losing ones into what he called team roles, each a cluster of behavior that helps a team make progress. The Belbin company now lists nine: Plant, Monitor Evaluator and Specialist for thinking; Shaper, Implementer and Completer Finisher for getting the task done; Resource Investigator, Teamworker and Co-ordinator for working with people. The Specialist was added after the Henley research, once Belbin applied it in industry.
Each role comes with what Belbin called an allowable weakness. The Plant, full of unorthodox ideas, may be forgetful. The Completer Finisher, who polishes and checks, can take perfectionism too far. The point was balance. By the company’s account, a team with no Plant struggled for a first idea, and a team with too many let bad ideas hide good ones. With too many Shapers, the drivers, the infighting began. Read this way, the Apollo companies had plenty of one kind of contribution and too little of the others.
What the questionnaire can and can’t tell you
After the research, Belbin turned the roles into a questionnaire, the Self-Perception Inventory, along with an assessment filled in by colleagues. Here the evidence is more mixed than the framework’s popularity suggests.
In 1993 the psychologists Adrian Furnham, Howard Steele and David Pendleton published a psychometric assessment of the self-perception inventory, and Belbin replied in the same journal. A 2007 review by Aitor Aritzeta, Stephen Swailes and Barbara Senior covered 43 empirical studies. They called the evidence mixed and concluded that, on balance, the model and inventory showed adequate convergent validity, but that some scales were so strongly associated with each other that they were hard to tell apart. A 2008 study of a revised, multi-rater version, with at least four observers for each of 1,434 people, found its validity good.
Whether role balance makes teams win is a separate question. A 2013 study of 24 student teams in a management game found no relationship between role diversity and performance, measured by the teams’ report grades and their financial results. When Google studied 180 of its own teams, what mattered, by the company’s account, was less who was on a team than how it worked together, though Google noted that factors that didn’t register in its data might still matter elsewhere. The Apollo result itself comes from Belbin’s account of his own games.
Four stars for one pilot
Imagine you’re picking a team for a three-month pilot that the whole company will watch. The obvious move is the four highest-rated people in the department, and that’s what you do.
By week three the meetings are lively. Every option gets taken apart. Two people have rewritten the plan and the other two have rewritten their rewrites. Nobody has called the five customers who agreed to test it, and nobody has written down what “done” means for the first release.
None of that needs a questionnaire to spot. It needs the list Belbin’s framework pushes you to write, of contributions rather than people: someone to go out and talk to the five customers, and someone to check the work and close it out. Next to both lines, you have no name.
A role isn’t a name tag
People who build teams of AI agents reach for the same vocabulary. One agent is “the reviewer,” another “the architect.”
In a study in Findings of EMNLP 2024, Mingqian Zheng and colleagues tested 162 personas, from “lawyer” to “AI assistant,” on 2,410 multiple-choice questions, using open-weight models from four families, including Llama 3, Mistral and Qwen2.5. Adding a persona such as “You are a lawyer” to the system prompt did not improve accuracy compared with no persona. A persona may help in certain settings, the authors wrote, but the effect of each one “can be largely random.” They tested factual questions, not agent teams, and no closed models such as GPT-4.
Belbin’s roles were behavior, not titles. For an agent, the behavior lives in the job: what it is allowed to read and change, and the check it has to run. A Completer Finisher is a step that runs the test suite and lists every failure before anything ships. “You are meticulous” is a sentence. Run your setup with and without the role line, a few times each, and judge the outputs against the task’s own standard. If they barely move, the word at the top of the prompt was a name tag.