If AI and humans split work by what each is actually better at, which cuts matter — topic, job title, or something else?

This explores which ways of dividing a task let AI and people each take the parts they're better at. The collection's short answer: the split that works is rarely by topic or by job title. It runs along a few cuts that show up again and again. One is execution versus judgment. Another is a short time budget versus a long one. A third is who decides what the task is and how to do it.

The clearest cut is between executing and deciding. Narayanan and Kapoor argue that AI shrinks only the middle 'execute' layer of knowledge work. Deciding what to do and delivering it to a real client stay with people, and those layers can even grow, which helps explain why translation and legal jobs haven't disappeared Does AI really compress all layers of knowledge work equally?. A study of 769 tasks from building a real research system found the same split. Agents proposed methods and carried out revisions, while humans made most of the final calls and chose which directions to explore. Participants said about a third of those tasks wouldn't have been feasible without the agent How should AI agents and humans divide research tasks?. Socher's account of reward hacking explains why the judgment layer stays human. AI is good at satisfying what was literally asked and weaker at grasping what was meant, so someone has to hold the intent Why do AIs keep gaming rewards instead of serving intent?.

The less obvious cut is time. On METR's research-engineering benchmark, AI agents scored 4× higher than human experts when both had two hours. Humans pulled roughly even at eight hours and led 2× at 32 hours When do AI agents outperform human research experts?. So comparative advantage depends partly on the clock. Short, well-defined bursts favor agents, and long work that builds on itself favors people. That suggests a practical structure: break long projects into short sprints for the agent, with a human holding the thread between them. Real usage data points the same way. Workers have actually handed off tasks in information-heavy jobs, and the pattern follows what the technology can reliably do more than which jobs are 'routine' Where have workers actually delegated tasks to AI?.

The same task can also be split in different ways depending on who holds control. Among 244 BCG consultants, three styles emerged. 'Centaurs' handed off whole chunks of work, 'Cyborgs' wove AI through every step, and 'Self-Automators' let AI take over. What separated them was who chose the task and who chose the method, not the tool or the task itself Do humans and AI divide labor differently in knowledge work?. The split also changes who benefits. When AI use was required, lower-performing consultants gained the most. The authors warn that this leveling effect can reverse when people choose for themselves whether to use AI Does AI help lower-skilled workers catch up to top performers?.

The twist you might not expect is that a good split doesn't necessarily save time. It moves time around. Hours that went into doing the task shift toward writing prompts and checking AI output Does AI really save time, or just change how we spend it?. That means the human side of any division of labor is mostly evaluation, which is a different skill from doing the work yourself. Trust also builds with repeated interaction. In repeated partner-choice games, people started out biased against AI partners but came to prefer them because the bots behaved more consistently Do humans learn to prefer AI partners over time?. If you want to go further, the open question the collection doesn't settle is whether people stay good at judging AI output once AI does most of the doing.