Human-AI Collaboration Is a Question of Who Hands Work to Whom

Almost every experiment puts the person last, deciding after the AI speaks. Workers, experts and one study that let the model hand out the work suggest that the direction of delegation is a design choice.

4 min read AI & Agentic SystemsIS Theory

In an earlier note I argued that delegation, not use, is the better lens for agentic systems. This note asks what the human-AI collaboration evidence says about the direction of that delegation. The short answer is that almost every study holds it fixed.

Baird and Maruping (2021) describe delegation as "transferring rights and responsibilities for task execution and outcomes to another". They also argue that "Agentic IS artifacts are no longer passive tools waiting to be used, are no longer always subordinate to the human agent". If that is right, then who hands work to whom should be one of the main variables in research on human-AI collaboration.

In practice it mostly is not. Vaccaro, Almaatouq and Malone (2024) report that "Most (>95%) of the human-AI systems in our dataset involved humans making the final decisions after receiving input from AI algorithms." They also found that "Only 3 of the 100+ experiments in our analysis explore such processes with a predetermined delegation of separate subtasks to humans and AI." So the field's picture of human-AI collaboration rests very largely on one arrangement. The model advises, and the person decides.

Hemmer et al. (2023) tried another arrangement. In their experiment the model decided which images each person should classify, so that "Humans receive exactly those images that match their skills." The result surprised me. "In an experimental study with 196 participants, we show that task performance and task satisfaction improve through AI delegation, regardless of whether humans are aware of the delegation." The mechanism ran through people's belief in their own ability. "While self-efficacy partially mediates the effect of AI delegation on human task performance, it fully mediates the effect of AI delegation on task satisfaction." The authors are honest about the uneasy part. They ask whether "humans managed by AI models" can be a suitable form of collaboration for particular workplace settings.

Shao et al. (2025) asked the workers. Their study drew on "responses from 1,500 workers across 104 occupations and annotations from 52 AI experts, covering 844 occupational tasks", and it places each task on a Human Agency Scale from H1 to H5. At the middle level, H3, the "AI agent and human form equal partnership, outperforming either alone." They report that "45.2% of occupations have H3 (equal partnership) as the dominant worker-desired level". Workers and experts often disagreed, though. "Among 844 tasks, 26.9% receive matching levels from workers and AI experts." Where they differed, "workers generally prefer higher levels of human agency than what experts deem technologically necessary".

I read these pieces as one finding. The direction of delegation is a design variable. The lab has mostly held it fixed. And the people doing the work have views about it that do not match the experts' views. For IS this matters because the arrangement is set when the system is built, often by people who are not the workers. A model that hands out tasks changes who manages whom. A model that only advises leaves the person answerable for every final call, including the calls that the model alone would have made better. Neither arrangement is neutral.

What I would like to see in IS studies of AI at work is the delegation arrangement reported as carefully as the model's accuracy. Who assigns the task? Who can override? Who signs off, and did the workers have a say in any of it? Shao and colleagues add two cautions that I find fair. They write that "higher HAS levels are not inherently better", and they note that workers "may still lack full awareness of the evolving capabilities and limitations of AI agents". So neither the workers' preference nor the experts' judgement settles the question. That is exactly why the arrangement should be studied and not assumed.

There is a link here to the note on what each side knows. If the person is better on some cases and the model on others, the arrangement that sends each case to the stronger party should beat the arrangement where the person reviews everything. Hemmer and colleagues built something close to that, and Vaccaro and colleagues found almost no one else had tested it.

The limits are real. Hemmer and colleagues ran their study with non-experts on a task that needs no special training, and they say so. Shao and colleagues measured what workers want and what experts judge possible, not what happens when a given arrangement is put in place. I still think the pattern is clear enough to change how we design and report studies. When a paper says humans and AI collaborated, I now want to know who handed the work to whom.