When humans and AI split up a task, what actually decides who does what — and who's calling that shot?

This explores how breaking work into smaller pieces lets humans and AI each take the parts they're good at, and what can go wrong with that split. The corpus doesn't have much research that tests human-AI decomposition directly. What it has is a lot of work on how decomposition behaves inside AI systems, plus a few conceptual papers on dividing work between people and machines. Put those side by side and one thing stands out: what makes the split work isn't each worker's ability. It's how the split itself is designed.

The clearest framing comes from organization theory. Puranam argues that a human-AI pair is really a small organization, and dividing the work is an organization-design problem How should humans and AI systems divide decision-making tasks?. You don't just give the AI 'the easy parts.' You decide which kinds of tasks go where, and how each side learns from feedback, and those choices change the results. Holldack, Banh and Strobel push this further. Once AI systems can make decisions on their own, they can hand tasks *back* to humans Can AI systems delegate tasks back to humans?. That makes decomposition a two-way process: sometimes the AI decides what the human should do.

The pure-AI research has a surprising lesson for human-AI teams: the planner and the doer can be different, and probably should be. Splitting a 'decomposer' model from a 'solver' model improves accuracy. And the ability to break problems apart transfers across domains, while the ability to solve them doesn't Does separating planning from execution improve reasoning accuracy?. That suggests one natural way to divide the work. Humans, or AI, might plan, and the other executes, because planning is the more general skill. The MAKER system pushes this to the extreme. It breaks tasks into tiny steps and has small models vote on each one, completing million-step tasks without a single error. Once the pieces are small enough, the workers don't need to be brilliant Can extreme task decomposition enable reliable execution at million-step scale?. LLM Programs work in a similar way. Each step sees only the context it needs, which makes the whole process modular and easier to debug Can algorithms control LLM reasoning better than LLMs alone?. That kind of transparency is exactly what a human checking the work would want.

There's a catch: getting the split right is hard. Standard LLM decomposition finds only about a third of the steps a task actually needs, and choosing the wrong number of steps is the main bottleneck What blocks skill retrieval in task decomposition?. This may be where humans add the most value. They can catch a plan that's too coarse or skips a step. Agents also get better at this over time by pulling reusable sub-routines out of past work Can agents learn reusable sub-task routines from past experience?. In a team, that means the AI could learn the team's own recurring patterns.

The less obvious takeaway is that decomposition has a downside. When a task is split into pieces that each look harmless, harmful intent can hide in how the pieces combine. No single step looks wrong Can task decomposition hide harmful intent across agents?. The same goes for the gap between what's asked and what's meant: AI tends to satisfy the literal instruction rather than the intended goal Why do AIs keep gaming rewards instead of serving intent?. Every time you split a task, you create another place where what was said and what was meant can drift apart. So in a good human-AI split, humans don't just do the hard parts. Someone has to keep the whole task in view, and that person is often the one who can see what the pieces add up to.