Designing hybrid teams: what role should humans and AI play?

Many companies are introducing AI along a similar path. The first step is usually individual productivity: give everyone a Copilot licence or an equivalent tool, provide training, and ask people to incorporate it into their daily work. Each person looks for ways to save time: drafting a document, summarising a meeting, analysing data, or preparing a presentation.

The second step is to create agents that perform tasks with some degree of autonomy. They no longer wait for a one-off request; they follow instructions, consult information, produce an output, and in some cases chain together several actions. The company moves from helping people with AI to delegating parts of the work to AI systems.

The third step is more demanding: creating hybrid teams in which people and agents collaborate within the same workflow. At this point, simply distributing tasks is not enough. We need to decide who contributes ideas, who challenges a proposal, who turns decisions into actions, who reviews the outcome and, above all, who retains judgement and accountability.

This journey leads to a fourth step: rethinking the company’s processes. If we merely insert AI into existing processes, we may only do faster what we were already doing. Real change comes when we revisit how decisions are made, how information flows, and how humans and agents coordinate to create more value.

This article focuses on the third step. The question is no longer only what AI can do, but what role it should have within a specific team so that the whole team works better.

Belbin’s team-role model gives us a language for exploring this. I will not use it to label people or attribute a personality to AI, but to observe which contributions we already have, which are missing, and which we might strengthen.

First, understand what the team needs

Meredith Belbin described nine roles that tend to emerge when people work together:

• idea-oriented: Plant, Monitor Evaluator, and Specialist;
• action-oriented: Shaper, Implementer, and Completer Finisher;
• people-oriented: Coordinator, Teamworker, and Resource Investigator.

A person can display more than one role and adapt their contribution to the situation. The model works best as a tool for discussing strengths, gaps, and imbalances rather than as a fixed personality diagnosis.

This perspective changes how we integrate AI. Instead of starting by asking “what tasks can we automate?”, we can observe what is happening in the team. Perhaps it generates many ideas but struggles to prioritise them. Perhaps it makes decisions but fails to follow through. Or perhaps it analyses problems well but avoids difficult conversations.

Only then does it make sense to decide whether an AI agent could strengthen one of those contributions.

An experiment with ChatGPT, Claude, and Gemini

To explore the extent to which an AI can represent these contributions, I gave the same Belbin questionnaire to ChatGPT, Claude, and Gemini.

These were the results:

Belbin role	ChatGPT	Claude	Gemini
Coordinator	11	9	7
Teamworker	13	11	8
Resource Investigator	5	4	2
Shaper	3	3	0
Implementer	11	10	13
Completer Finisher	11	6	14
Plant	5	5	3
Monitor Evaluator	6	10	12
Specialist	5	12	11

The profiles are not identical. ChatGPT presents mainly as a Teamworker. Claude stands out as a Specialist, while Gemini concentrates on Completer Finisher, Implementer, and Monitor Evaluator.

There are similarities too. All three AIs score low as Shapers and Resource Investigators. None presents as especially strong in the Plant role, even though generating ideas is one of the most common uses of generative AI.

This last result surprised me. It also made me wary of reading too much into the test literally.

Simulating behaviour is not taking on a role

The Belbin questionnaire is designed for a person to assess their usual behaviour based on the experience of working with other people. An AI does not have that experience. It responds based on its training, the instructions, the conversational context, and its interpretation of each statement.

The results, then, do not prove that models have a team role. They show the behaviour they say they would display under specific conditions.

ChatGPT’s high score as a Teamworker is a good example. The model tends to respond kindly, adapt its tone, and avoid unnecessary conflict. These are behaviours associated with the role.

But a human Teamworker perceives silence, alliances, tiredness, and tension. They have built trust with other people and risk some of that trust when they intervene. AI can suggest an empathetic sentence, but it does not share the relational cost of saying it at the wrong moment.

Something similar happens with the Coordinator role. An AI can organise objectives, propose an agenda, and distribute tasks based on the available information. A person who coordinates also needs legitimacy, political awareness, and the ability to stand behind a decision in front of the group.

The visible behaviour may look similar. The full contribution is not the same.

AI’s role has to be designed

It would be a mistake to conclude that “Gemini is a Completer Finisher” or that “Claude is a Specialist”. The same model can behave very differently depending on how we configure it.

We can create one agent that generates non-obvious alternatives, another that looks for risks and contradictions, another that turns a decision into an executable plan, and a fourth that reviews the outcome before it is considered complete. All four can use the same model.

The difference does not come from a fictional personality. It comes from the objective, context, questions, criteria, constraints, and response format.

Giving five agents different names does not guarantee five different contributions either. If we only ask them to “act as” five roles, we may end up with the same model wearing five costumes. Each agent needs:

• a concrete function within the process;
• the information required to fulfil it;
• its own criteria for examining the problem;
• clear limits on what it can decide;
• an output the team can review;
• a person responsible for validating the result.

Designing the agent’s role matters more than assigning it a character.

What can AI contribute best?

The experiment points to contributions where today’s AIs can be useful: structuring information, turning a decision into steps, comparing options, looking for inconsistencies, checking details, and providing specialised knowledge when they have good sources.

This aligns with the Implementer, Completer Finisher, Monitor Evaluator, and Specialist scores, especially for Claude and Gemini.

A high score, however, does not guarantee good execution. A Completer Finisher agent can systematically review a text and still miss an invented fact. A Monitor Evaluator can compare options using incomplete criteria. A Specialist can give a precise answer based on a false premise.

That is why human oversight is not a formality added at the end. It is part of designing the hybrid team.

Contributions related to trust, conflict, influence, or social responsibility still depend much more on people. AI can support a Teamworker or a Coordinator, but it cannot replace the relationship that gives legitimacy to their intervention.

Adding AI does not always improve the team

When a tool is new, it is easy to introduce it wherever it looks most impressive. We ask for ideas, alternatives, documents, and analyses. Within minutes, we have more material than the team can review.

A team can have five agents generating ideas and still fail to innovate. Perhaps the problem is not a lack of ideas, but that nobody challenges them, prioritises them, or turns them into decisions.

AI can also intensify an existing imbalance. If the team already opens too many threads and struggles to close them, another alternative generator will only create more noise. In that case, an agent that compares options, makes decision criteria visible, or checks what is missing to finish would be more useful.

The unit of analysis should not be “humans versus AI”, but the whole team:

• What contributions are people already making?
• Which ones always fall to the same person?
• Which ones disappear under pressure?
• Where does the flow from an idea to an outcome get stuck?
• Which agent could strengthen a gap without weakening human accountability?

Belbin can help us formulate these questions. It will not give us an automatic recipe.

Five criteria for designing a hybrid team

I take five practical criteria from this experiment:

1. Start with the team, not the tool. Observe which contributions are missing before adding an agent.
2. Define a function, not a character. “Review the evidence and identify contradictions” is more useful than “act as a Monitor Evaluator”.
3. Keep judgement and accountability with people. An agent can prepare, compare, or review; a person must decide and answer for the outcome.
4. Design the review from the beginning. We need to know who will validate the output, using which criteria, and what will happen if there are doubts.
5. Measure the contribution to the whole. It is not enough for the agent to produce a lot or respond quickly. It must help the team make better decisions, reduce errors, or complete work that previously got stuck.

The conclusion

Designing a hybrid team is not about distributing tasks between people and AI until everyone is busy. It is about understanding what the team needs in order to work better and deciding which contributions technology can strengthen without diluting judgement, relationships, or human accountability.

The questionnaires do not prove that ChatGPT, Claude, or Gemini have a Belbin personality. They do help us see an important distinction: an AI can perform behaviours associated with a role without taking on that role as a person would.

The question left to us is more specific: what contribution does this team need, and how can we strengthen it with AI without delegating what must remain human?

In the second part, we will turn this question into a practical exercise for mapping the team’s contributions, identifying gaps, and designing AI agents with a verifiable function.