AI can give one person the output of a team. That changes management

A single capable person with AI can now draft the campaign, build the deck, run the first-pass analysis and write the supporting code, work that not long ago needed four people and a week. This is real, and it leads quickly to a conclusion that is half right and wholly dangerous: that you simply need fewer people. The more useful reading is that AI changes where a team adds value, not whether it does.

The trap is to treat output as if it were the same as judgement. AI is very good at producing a team's worth of output. It is not good at supplying a team's worth of challenge, scrutiny and diverse expertise, which is the part that stops confident work from being confidently wrong. Manage for the output alone and you get speed with no brakes.

What AI actually changes, and what it does not

The evidence points in a consistent direction. In the customer-support study by Brynjolfsson, Li and Raymond, the largest gains went to less experienced workers: AI raised the floor more than the ceiling, helping a capable generalist reach a competent draft in territory that used to require a specialist. That is exactly what lets one person cross boundaries that once needed a team.

But the same technology has a sharp edge. The "jagged frontier" field experiment found AI could make knowledge workers meaningfully worse on tasks just outside its reliable range, and confidently so. A solo worker plus AI has no one to catch that error; a team does. So AI does not remove the need for expertise and challenge. It raises the stakes on them, because the work arrives faster, looks more finished, and has had fewer eyes on it. The MIT Media Lab's GenAI Divide report makes the related point that value comes from integration and judgement, not from output volume.

Three modes of AI-supported work

The management move is to stop asking "can one person do this now?" and start asking "which mode does this work actually need?" There are three.

Solo plus AI. One person and AI handle the whole task. Right for self-contained, lower-stakes, reversible work where speed matters and a wrong answer is cheap to catch and fix. This is most day-to-day work, and AI makes it genuinely faster.

Expert plus AI. An expert works with AI on something where quality and judgement matter. The expert is not there to type faster; they are there to supply the discernment AI lacks, to know when a fluent answer is wrong. Right for high-stakes individual work where one wrong call is costly.

Team plus AI. A team works with AI on work that is genuinely cross-disciplinary, contested, or high-consequence. Here AI amplifies each member, but the team is doing something AI cannot: challenging each other, bringing real different expertise, and catching what an agreeable model and a single tired human would miss together.

One person plus AI can produce a team's output, not a team's judgement.

What this looks like in practice

Take a marketing agency. A consultant with AI can turn around a campaign draft or a client report quickly and well: solo plus AI, and a real gain. The strategy for a major client moving into a new market is a different mode entirely. That work is contested, high-consequence and cross-disciplinary, and it needs the challenge of a team plus AI. Run it as solo plus AI because the tool makes it possible, and you get something polished, plausible and unchallenged, which is the most expensive kind of wrong. (An illustrative example, not a specific agency.) The failure is not using AI. It is assigning the wrong mode to the work.

There is a management point underneath all this, and it is easy to miss. As AI lets work be restructured, with fewer people producing more, the temptation is to let accountability thin out along with the headcount. It must not. The OECD's AI Principles keep human oversight and accountability central precisely because efficiency makes them the first things to quietly drop, and a one-person-plus-AI output still needs a named human who owns whether it is right and answers for it if it is not. Choosing the mode is therefore also a decision about who is accountable, not only about who is fast. The more finished the work looks, the more that matters, because speed and polish are exactly what stop people asking whether anyone actually checked it. A team that has quietly become one person and a model has not removed the need for an owner; it has concentrated it.

The honest limits

Two cautions. First, this is not a claim that AI changes nothing about headcount. It genuinely lets smaller teams do more, and that can be a real commercial advantage; pretending otherwise helps no one. The point is that the saved capacity is best spent on more judgement and challenge per piece of work, not treated as proof that judgement and challenge are now optional.

Second, the modes are not fixed labels. A task can start as team plus AI while it is novel and contested, then settle into expert or solo plus AI once the judgement is well understood and the risks are known. Reassess as the work matures rather than assigning a mode once and forgetting it.

What to do next

For each significant piece of work, choose the mode on purpose. Ask how reversible it is, how contested, and how much it would cost to be confidently wrong. Let the cheap, reversible work run solo plus AI and enjoy the speed. Reserve expert and team modes for the work where a polished wrong answer would actually hurt, and resist the pull to default everything to solo plus AI simply because the tool now makes it possible.

The tool

To make that call quickly, we have built the AI Work Mode Selector: a short decision guide that takes a piece of work and points you to solo, expert or team plus AI, based on its stakes, how contested it is, and how easily a mistake would be caught.

Download the AI Work Mode Selector (PDF)

Choosing modes well, and building the team habits around them, is part of the practical work of an AI Advisory engagement, and the judgement behind it is what our AI lessons for leaders build. It connects to why experts can be slowed by AI and to the decisions that should never leave a human.

Sources and further reading

• Brynjolfsson, Li and Raymond, Generative AI at Work, NBER Working Paper 31161, 2023. Independent. Source for AI raising the floor more than the ceiling, with the largest gains for less experienced workers.
• Dell'Acqua et al., Navigating the Jagged Technological Frontier, Harvard Business School / BCG working paper, 2023. Independent field experiment. Source for AI making workers confidently worse on out-of-frontier tasks.
• MIT Media Lab (Project NANDA), The GenAI Divide: State of AI in Business 2025. Industry report, not peer-reviewed; the headline 95% figure has been contested. Source for value coming from integration and judgement, not output volume.
• OECD AI Principles. Adopted 2019, updated 2024. Source for human oversight and accountability remaining central as work is restructured.