Should you give each task to whichever is better, you or the AI — or does that simple rule break down in practice?

This explores how organizations should split work between people and AI, and whether 'give each task to whoever does it best' actually holds up. The corpus suggests that sorting tasks one at a time is the wrong starting point. How should humans and AI systems divide decision-making tasks? treats a human-AI pair as a small organization. The useful questions are what kind of task it is, how feedback reaches whoever is learning, and how the final decision gets made. The same split of work can succeed or fail depending on those settings.

The idea of 'comparative advantage' also turns out to be less stable than it sounds. In METR's research-engineering tests, AI agents scored about 4× higher than human experts when both had two hours. With eight hours, the humans were slightly ahead, and with 32 hours they led by about 2× When do AI agents outperform human research experts?. So who has the advantage can depend on the time budget as much as on the task. Fast, bounded work suits agents. Long work that improves with sustained effort still favors people. On the real-world side, the tasks workers have actually handed to AI cluster in information-heavy jobs and follow what the technology can do, not the older predictions about 'routine' work being automated first Where have workers actually delegated tasks to AI?.

The study of 244 BCG consultants adds a useful twist. People with the same tool and the same task divided the work in three different ways: Cyborgs blended closely with the AI, Centaurs handed off clear chunks, and Self-Automators let the AI do most of it. What decided the pattern was who chose the task and who chose the method Do humans and AI divide labor differently in knowledge work?. Allocation is partly about control, not just skill. Handoffs can also run the other way: agentic systems may start assigning tasks to humans Can AI systems delegate tasks back to humans?. For agent workflows where nobody can know the 'right' moment to ask a human, Magentic-UI doesn't try to solve that timing problem. It spreads human checkpoints across co-planning, action guards and verification When should human-agent systems ask for human help?.

The part most task-by-task plans miss is how decisions depend on each other. Does AI adoption require rethinking how organizations coordinate? models a case where AI makes each individual decision better but also makes decisions vary more. When decisions depend on each other, that extra variation creates coordination problems that cancel out the local gains unless the organization redesigns how its parts communicate. Agentic AI also fits awkwardly into existing management categories. Most executives surveyed see it as a coworker, yet it needs to be managed partly like staff and partly like equipment Does agentic AI fit the tool-or-coworker framework?. Recognizing which tasks could be handed off at all is another bottleneck, because most workers don't see their own work as automatable Does easier tool-building actually solve enterprise adoption problems?.

Finally, the split is not only an efficiency question. Brynjolfsson argues that choosing to automate rather than augment moves bargaining power toward whoever controls the technology. That can lock in a situation where workers lose the leverage they would need to improve outcomes Does automating labor concentrate power more than augmenting it?. The takeaway is that 'who's better at this task?' is a starting point, not the answer. Allocation decisions also determine who keeps learning, who holds control, and how well the whole organization fits together.