AI Never Made It Into The Room

Your adoption numbers are probably fine. Most of your managers use AI several times a day, and the dashboard shows it. Almost none of them use it in the room where the work is actually decided.

Every hand went down

Gabriele Rosani and Elisa Farri describe a town hall they ran with 200 senior managers and executives of a large energy company, in a piece for Harvard Business Review this month. They asked the room to raise a hand if they use AI multiple times a day at work. Every hand went up. Keep it up if you also use AI together with your teams and colleagues. Nearly every hand went down. Raise it again if your days are packed with meetings, calls and work sessions with those same teams. All hands up.

Their line for it: AI use "was contained to one person, one screen, one task." The same pattern held across the more than 300 managers in 35 organisations they studied. Their conclusion is blunt: "most organizations overlook where much of their work happens: in teams."

The numbers agree. Myles Suer, reviewing their new HBR Guide in InformationWeek, cites Capgemini data: daily generative AI use among managers went from 15% in a 2024 survey of 1,500 managers to 49%. Only 7% use it for team activities. Team use barely registers.

Why the dashboard cannot see it

I understand why nobody has noticed. Adoption is measured per person: licences, active users, prompts per week. Every one of those counts individuals, so a company where every manager uses AI alone and nobody uses it together scores a perfect hundred. The instrument is built to miss the gap.

Where AI has entered the meeting room, it usually came in as a transcription bot. It records, it summarises, it sends the action list. Suer's phrase for that is "a glorified stenographer", set against the alternative of AI as "an active member of the team". The bot is the alibi. It lets an organisation say AI is in its meetings while nothing about how the meeting thinks has changed. The proposal is presented the way it was in 2022, questioned the way it was in 2022, and decided the way it was in 2022. Only the minutes arrive faster.

Ethan Mollick wrote last week about the overhang: the "capability overhang between what those models can do today and what most folks are using them for is massive", and the bottleneck is "how slowly our very human systems and processes work to keep up". Read that against the show of hands. A large part of that overhang sits in the meeting room, because that is where the human systems live and where the tool has not been let in.

Only one person holds the key

Rosani and Farri name two conditions for AI in teamwork. The second is craft: structure, the right context, good prompts, and the team keeping collective ownership by shaping prompts together and debating the output. That is teachable. The first is intentionality, and their wording matters: "only the team leader has the agency to decide to change how the team activity is conducted."

Nobody else can do this. The enthusiastic analyst cannot open a model on the shared screen in someone else's meeting. L&D cannot train its way into a Tuesday planning session it does not run. The person who owns the agenda decides whether the team works with the tool or only next to it. So the capability we have been building, one person at a time, has been built for the wrong unit. Decisions are made in teams, and most training programmes still certify individuals.

The fair objection is that meetings are for people, and keeping AI out of them is a feature. I have some sympathy. A room staring at a model's output is not a better room. But that is an argument about how, not whether. Rosani and Farri's research places the useful moments precisely: before the meeting, to sharpen preparation and pressure-test ideas; during it, to challenge assumptions and broaden perspectives; after it, to reflect and turn lessons into reusable practice. None of that replaces the conversation. It gives the conversation something harder to push against.

One meeting, this week

If you run a team, you do not need a budget for this. Pick one recurring meeting where a decision gets made. Before the decision, put the model on the shared screen and ask it to argue against the proposal. Write the prompt together, out loud, so the team owns the question as much as the answer. Then argue with what comes back. Take five minutes at the end to ask what it changed, if anything. Save the prompt where the team can find it, and use it again next week.

It costs nothing, and it moves the tool from the private screen into the ritual where the organisation actually decides things. If you work in L&D or transformation, you will not be the one running it. Your job is to find the team leads who will, and make it easy for them. That is what we build with people in the AI Capability Academy programme, but the first version needs nothing more than a screen and a meeting you already have.

You measured adoption and got a good number. Now count the rooms.