What happens when the whole team starts using AI?

Generative AI has become personal at work. We sit at our own screens asking it to draft an email, analyse information, challenge an idea or help us work through a problem. Yet much of the work organisations rely on happens with other people. Look at most senior leaders' calendars and a significant part of the week is spent discussing work with colleagues rather than producing it alone

This mismatch prompted Gabriele Rosani and Elisa Farri of Capgemini Invent's Management Lab to investigate what happens when generative AI becomes part of the group rather than something each employee uses alone.

Farri describes the starting point rather neatly. “Nobody asked us how to use generative AI with other people, with their teams, in collective settings,” she said on a recent Harvard Business Review webinar. 

Their subsequent research suggests this remains unusual. Around half of leaders and managers are using generative AI individually at work according to figures they presented during the webinar while fewer than 10% are actively integrating it into group settings. Their new HBR Guide to Generative AI for Teams explores how AI might be used in activities ranging from strategy workshops and project reviews to problem solving and decision making. 

There is an interesting idea running through their work. Using AI with other people may help to counter some of the behaviours that can develop when we use it by ourselves.

Using AI together can slow us down in a good way

Rosani's concern is what can happen when the exchange becomes too frictionless. Working alone with AI makes it easy to accept a convincing response and move straight to the next question without spending much time examining how the answer was reached. Put colleagues around the same response and the pace naturally changes because people interpret it differently and may want to pursue parts of it that somebody working alone would simply have accepted.

“When you are with other people, of course you slow down a little bit,” he says. He sees this as part of the value of collective use because the discussion gives people more opportunity to exercise their own judgement before deciding what to do with the AI's contribution.

In experiments involving more than 300 managers Rosani says around two-thirds reported higher-quality outcomes from working with AI collectively. A similar proportion felt some of the risks associated with AI had been mitigated through what he calls “collective judgement”. 

These are findings from their own experiments rather than evidence that adding AI will automatically improve a meeting. Even so, they widen the question organisations need to ask about AI capability. Teaching an employee to use a chatbot effectively is only part of the picture if the decisions they make at work are usually shaped with other people.

Are we as a team owning the conversation or is AI leading it?
Gabriele Rosani, co-author, HBR Guide to Generative AI for Teams

Ask AI to challenge the team

Some of the strongest examples from Rosani and Farri involve giving AI permission to disagree. In one innovation workshop a team brought AI into the discussion after it had already developed an initial business concept. One agent was asked to challenge the thinking while another responded from the perspective of a customer. Crucially, neither was designed simply to deliver a verdict. The AI raised questions and then waited while the people in the room worked through them before the conversation continued.

Rosani describes AI in this setting as a “sparring partner”. The value comes from what the team does in response. If AI simply produces a critique and everyone reads it very little has changed in the way the group thinks together.

Farri saw the same principle applied during the quarterly review meeting at a luxury company'. Participants used AI before the meeting to question the recommendations they intended to bring. When they later worked in small groups AI adopted the perspective of a competitor CEO and pushed back on their ideas again. The purpose throughout was to give the team something worth arguing with rather than something to follow.

Use AI before the meeting starts

This example also shows that AI does not have to be running visibly throughout a meeting to influence its quality. Participants used AI before the meeting to put some pressure on the recommendations they planned to make. By the time they spoke to colleagues their thinking had already been questioned once and the meeting could start further into the discussion.

This may be a sensible route for teams that are less confident with generative AI. Rosani suggests that where a team is not yet ready to use it actively during a session, leaders can begin by using AI to prepare for the meeting and become familiar with the interaction first. 

There is basic meeting benefit here as well. People often arrive with an idea they have barely had time to test and everybody else then spends part of the meeting doing that testing with them. A short conversation with AI beforehand may help someone discover where their argument is thin before taking up the team's time.

Do not let the meeting become people watching AI

There is also an obvious way for all this to go wrong. Someone opens the AI tool, enters a prompt and everybody else watches the answer appear. Interactions must create a genuine exchange between AI and the group. “If it's not designed this way... you're going to kill teamwork,” Farri says. “People will switch off. They will just watch a screen.” 

Farri's answer is to make participation part of the way the AI interaction is designed. In some sessions the tool explicitly asks people for their views and waits while the group talks. In another experiment one person handled the typing while the rest of the team concentrated on the discussion, which avoids turning prompting skill into the centre of the exercise

This puts quite a lot back on the person running the meeting. They need to know why AI is there in the first place and be willing to stop using it when the conversation between colleagues is producing more value. Rosani suggests one question: “Are we as a team owning the conversation or is AI leading it?” 

Start with 10 minutes

There is little reason to turn the next leadership meeting into an elaborate AI workshop. For teams beginning to experiment Rosani suggests using AI as a challenger for perhaps 10 minutes at the end of an existing session. The temptation is to upload the group's work and ask for a list of weaknesses. He warns that this simply produces another AI answer for people to consume.

A better approach is to make the challenge conversational. Ask AI to question the team's thinking, let people respond and allow the next question to build on what they say. 

Organisations have spent the past few years encouraging employees to find individual uses for generative AI. Rosani and Farri are asking leaders to take the next experiment into the meeting room and pay close attention to what happens there.

Their work also provides a warning against assuming that simply adding AI makes a team more capable. If people stop debating because the machine has produced something plausible then the quality of the discussion is likely to suffer. AI has more value when its contribution prompts colleagues to keep thinking and talking rather than bringing the conversation to an early close.