In human-AI teamwork, nobody actually knows the rule for when the AI should hand off to a person instead of pushing ahead.

This explores why deciding who handles a given piece of work (the AI, a human, or another agent) matters so much when humans and AI share tasks. The corpus partly supports the question's premise, with one important change. Routing is central, but it is rarely one decision made up front. It works better as a series of small handoffs that keep happening throughout a task. Puranam's framing is the clearest starting point. A human-AI pair is a small organization, and splitting the work is an organization design problem. The right split depends on the type of task and on how each side gets feedback and learns from it How should humans and AI systems divide decision-making tasks?.

The surprising part is that nobody has a reliable answer to 'when should the AI hand off to a human?' The Magentic-UI work says plainly that there is no ground truth for when deferring is best. Instead of trying to solve that, the designers spread the decision across six touchpoints: planning together, working on tasks together, guards on risky actions, verification, memory, and multitasking When should human-agent systems ask for human help?. Cocoa takes a similar approach at the level of individual plan steps. Users mark each step as theirs or the agent's and can reassign steps mid-task, which gave them more control without adding effort Can users steer AI agents better by revising plans mid-execution?. When routing can't be computed in advance, the fix is to make it easy to change along the way.

The direction of the handoff is also changing. Holldack, Banh and Strobel argue that agentic systems can now hand tasks *back* to humans, so delegation runs both ways and older one-way theories no longer fit Can AI systems delegate tasks back to humans?. Conversation analysis offers a precise way to decide when an agent should stop and ask the user. It calls these moments 'insert-expansions': clarifying questions that prevent the drift that sets in when an agent quietly chains tool calls on its own When should AI agents ask users instead of just searching?. Underneath this sits a deeper problem. Models have no internal record of what they don't know about the user. Adding an explicit list of unknowns to the prompt cut sycophancy and hallucination roughly in half Do language models know what they don't know about users?. An AI can't know when to ask until it can tell what it's missing.

On the human side, the trigger for taking work back isn't what you might expect. People didn't withdraw trust because the stakes were high. They withdrew it when an action was *irreversible and visible to others*, like sending an email, even when the AI's output was fine What makes people distrust AI agents they delegate to?. That gives routing a practical rule: send reversible work to the AI and keep a human checkpoint on anything that can't be undone. Some researchers also question whether the choice has to be 'AI decides or human decides' at all. 'Learning to guide' has the AI highlight what matters in the input while the human makes the call. That removes the anchoring bias that pure deferral creates Can AI guidance reduce anchoring bias better than AI decisions?.

Routing also comes up in two other settings. Among agents, versioned capability vectors turn 'who can do this?' into a searchable lookup instead of hand-wired connections Can semantic capability vectors replace manual agent routing?. In the real economy, workers have actually handed AI tasks that follow what the AI can technically do, mostly information-heavy work. Their own habits with chat-based AI turned out to be a weaker guide Where have workers actually delegated tasks to AI?. The takeaway: knowing who to ask matters because the cost of a bad handoff shows up later as drift, misplaced trust, or a mistake that can't be undone. The best systems handle this by keeping the question 'who should do this?' open for the whole task.