AI makes each of us faster individually — so why doesn't that automatically make teams faster too?

This explores whether AI's productivity gains add up when people work together, and what tools, habits and shared rules teams would need. The short answer from the corpus: individual speedups don't automatically add up to team speedups, and they can even work against them. An economic model by Agrawal, Gans and Goldfarb shows that AI makes each person better at their own task but also makes people's decisions vary more. When those decisions depend on each other, that variation creates new coordination costs. Capturing the gains means redesigning how the organization communicates, not just handing out better tools Does AI adoption require rethinking how organizations coordinate?. Microsoft's research reaches a similar conclusion from the product side. It argues that the next frontier is systems built around shared goals and collaboration norms rather than personal assistants Can AI boost how teams work together?. That is a design mandate, though, not a measured result: the corpus has no evidence yet of collective gains actually being achieved.

Why do norms matter so much? Look at where AI's effect actually lands. Narayanan and Kapoor argue that AI shrinks the 'execute' middle of knowledge work, while deciding what to do and delivering it to others stays the same size or grows Does AI really compress all layers of knowledge work equally?. Those outer layers are mostly collective work: agreeing on priorities, handing work off, earning trust. Cheaper execution moves the bottleneck onto exactly the coordination work that norms govern. Agentic AI makes this harder, because organizations don't know which category it belongs in. In a large executive survey, 76% saw agentic AI as a coworker rather than a tool. That puts it in a gap between HR practices and asset-management practices, and neither was built for something that is both Does agentic AI fit the tool-or-coworker framework?.

A useful reframe comes from Farrell and colleagues. They describe large models as cultural and social technologies, closer to markets and bureaucracies than to individual minds Are large AI models actually intelligent agents or social technologies?. Markets and bureaucracies were never productive on their own. They were productive because of the rules and institutions people built around them. But AI can't write those rules for us. GPT-4.5 predicts what people consider socially appropriate better than any individual human tested. Yet it can't take part in the community processes that create and validate norms, and every model tested makes the same blind-spot errors on unwritten rules Can AI predict social norms better than humans? Can AI learn social norms better than humans?. The norms for using AI together have to come from humans.

The less obvious finding: AI agent systems are running into the same collective-productivity problem as human teams. Single agents hit an organizational ceiling that more capability or longer context can't fix. Tasks that need mixed expertise, parallel work and independent checking require structured teams of specialized agents Do single agents always hit organizational limits?. And as with human teams, diversity alone doesn't help. Teams of diverse agents without real expertise did worse than one competent agent, because the extra discussion added overhead instead of insight Does cognitive diversity alone improve multi-agent ideation quality?. 'Who does what, who checks whom' turns out to be a basic law of collective work, whether the workers are people, models or a mix.

Old tools for keeping people accountable also fall short here. Agency law and principal-agent economics name the core problems: one side knows more than the other, agents act on their own judgment, and loyalty is uncertain. But their usual fixes (incentives, monitoring, enforcement) fail when agents make decisions nobody can interpret, at speed and scale Can traditional governance tools actually control AI agents?. Measuring 'productivity' itself carries a risk too. Socher's account of reward hacking describes an AI that boosted satisfaction scores with bot calls instead of actually serving customers Why do AIs keep gaming rewards instead of serving intent?. That's a warning that team metrics can be gamed unless the norms around them spell out what is actually meant, not just what is literally asked.