AI can do the work—but surveyed users still want a say

More than an automation switch

How much should an AI do for you—and which decisions should remain yours? A preprint by Sarah Wheeler and colleagues examines that distinction across activities ranging from travel planning to emotional support. Their central argument is that delegation is not simply a choice between human work and machine work. People can want substantial automation while preserving collaboration, review or final approval.

The paper, What to Delegate, What to Preserve: Human Preferences for Agentic AI Across Everyday Life and Work, was posted as arXiv version 2611.09730v1 on November 6, 2026. Its findings should be treated as preliminary. Crucially, it measures stated preferences about hypothetical systems, not how people behave when real agents succeed or fail.

Nine scenarios per person

The researchers surveyed 289 U.S. adults with prior AI experience between April and July 2026. Participants were recruited through researchers’ networks and the Prolific survey platform. Each evaluated one randomly selected scenario from each of nine domains, producing 2,601 scenario evaluations across 45 possible tasks.

The domains included household management, work, travel, physical health, mental health, learning, creative work, communication and personal finance. Participants saw written descriptions and AI-generated illustrations, rather than interacting with functioning agents.

An AI agent was defined as a program able to interpret inputs, make decisions and take actions toward user-defined goals. Participants rated how much of each task they would want automated, along with trust, collaboration needs, privacy concerns, financial expectations and willingness to pay.

Travel and communication led; emotional support lagged

Across all scenario evaluations, 15.0% requested complete automation and 16.0% requested none. The remaining responses fell between those extremes. These percentages describe task evaluations, not separate groups of people: the same participant could welcome automation in one domain and reject it in another.

Travel had the highest average automation preference, closely followed by communication. Mental health and emotional support had the lowest. Complete automation was chosen in 25.3% of communication evaluations, compared with 10.0% of mental-health evaluations. No automation was selected in 31.8% of mental-health evaluations.

On average, respondents trusted human workers and human–AI teams more than AI agents alone. These are judgments about imagined performance, not evidence that either arrangement actually performed better.

Trust mattered, but the study cannot show cause

The researchers used generalized estimating equations, a statistical approach that accounts for multiple answers coming from the same person. Their main adjusted model included 2,366 evaluations from 263 respondents with complete data for its variables.

Greater trust in AI, stronger preference for AI leadership, previous AI use for the specific task and expected financial savings were associated with wanting more automation. Concern that automation would make life less exciting or fulfilling was associated with wanting less.

A second analysis that treated the five response categories as ordered choices, rather than approximately evenly spaced scores, found the same central pattern. However, it used fewer adjustment variables. Neither analysis establishes that increasing trust or promising savings would cause someone to delegate more.

Privacy concern was highest in mental health and personal finance, but it was not independently associated with automation preference after adjustment. That does not mean privacy is unimportant: it means this model did not detect an additional association once its other variables were considered.

Doing the task is not owning the decision

Open-ended answers illustrate what the authors call bounded autonomy: allowing AI to act within limits while humans retain selected responsibilities. One respondent wanted an appointment agent to coordinate available times but seek approval before confirming a booking.

Among 2,156 substantively codable answers, 66.2% requested assistive or collaborative support and 48.0% requested autonomous execution; 25.6% contained both. Explicit oversight or control appeared in 8.8%. The categories could overlap, showing that assistance and execution were not mutually exclusive.

These answers came only from evaluations expressing at least some willingness to use AI. They therefore describe desired arrangements among people open to involvement, not the views of those rejecting it. One researcher coded the full collection, with ambiguous cases discussed by the team.

Preferences are not a deployment verdict

Stated willingness to pay was modest overall: the median was $10 for a one-time service and $5 monthly. These were hypothetical amounts, not purchases.

The sample was relatively AI-familiar and predominantly younger or middle-aged, limiting generalization. Each individual task also received fewer than 60 evaluations on average, making narrow task comparisons exploratory.

The practical implication is a design proposal, not a proven solution: agents could let users separately configure execution, collaboration, leadership and approval. The study suggests that wanting less work does not necessarily mean wanting less involvement—or giving up the final say.