Working with AI, people naturally fall into different styles of control — why does the same tool make one person a co-pilot and another a delegator?

This explores how people working with generative AI end up splitting the work. The corpus treats Cyborgs, Centaurs, and Self-Automators as three working styles that emerge from who controls what, not as roles a team assigns on purpose. The key study followed 244 BCG consultants and found that the three styles split along two questions: who picks the task, and who picks the method for doing it Do humans and AI divide labor differently in knowledge work?. The surprise is what didn't decide it. The same tool on the same task produced different styles in different people. So the split isn't something the technology imposes. It comes from where people choose to keep control.

That makes the topic a question of organization design. How should humans and AI systems divide decision-making tasks? treats a human paired with an AI as a small organization with two agents. The design choices are how to divide tasks and how each side learns from feedback. Seen this way, the three styles are three different organization charts. Puranam's main point is that these setups don't just feel different; they lead to different results depending on how feedback flows. In practice, choosing a working style also means choosing what you and the AI will be able to learn.

Individuals often don't get to choose freely. A study of junior and senior engineers found that company rules, such as required tools, approved-tool lists, and data policies, set how much control engineers keep before personal preference comes into play Does personal preference shape how engineers use AI tools?. And the split can also run the other way. Can AI systems delegate tasks back to humans? argues that agentic systems can hand tasks back to humans, which none of the three styles accounts for.

Here's the twist you might not expect. Who formally holds authority and who feels in control are separate things. Does attributing agency to AI actually reduce human sense of control? shows that when people credit AI with judgment of its own, their sense of control and responsibility drops even though they still have formal authority. This happens in a cycle of delegation, attribution, contingency, reassertion, and reconfiguration, so it can be reversed. A Self-Automator may therefore give up more than the task. They may give up their sense of owning the result. Design work on agent systems points the same way. Magentic-UI doesn't try to solve when the agent should defer to a human. It spreads human decision points across six mechanisms: co-planning, co-tasking, action guards, verification, memory, and multitasking When should human-agent systems ask for human help?. Data & Society researchers warn that approval checkpoints alone aren't real oversight if people can't see what the agent did or why Does human-in-the-loop design actually enable effective oversight?.

The corpus is thin on one point. It names the three styles and what produces them, but it doesn't show that any one style reliably beats the others, or how a team should match a style to a task. The closest it comes is the argument that human-AI collaboration outperforms full automation in research Can human-AI research teams improve faster than autonomous AI systems?. That is a reason to be careful about drifting into Self-Automator mode by default.