What to Delegate, What to Preserve: Human Preferences for Agentic AI Across Everyday Life and Work
Authors: Sarah Wheeler, Veronica Rammouz, Yaoyao Qian, Zexuan Liu, Mengyi Wei, Anthony Rios, Tianzhi He, Kyrie Zhou
Organizations: University of Illinois Urbana-Champaign, USA · University of Texas at San Antonio, USA · Northeastern University, USA · Technical University of Munich, Germany
AI agents can execute multi-step tasks on users' behalf, yet the boundaries of acceptable delegation remain unclear. We report a scenario-based survey of 289 U.S. adults with prior AI experience covering 45 use cases across nine life domains. Each respondent evaluated one scenario per domain, producing repeated measures of automation preference, trust, collaboration, AI leadership, privacy concern, and willingness to pay. Automation preferences varied substantially across domains. Generalized estimating equation models showed that greater trust in the AI agent, desired AI leadership, prior task-specific AI use, and anticipated financial savings were associated with greater automation preference. In contrast, concerns that automation would reduce personal fulfillment predicted lower automation preference. Open-ended responses further revealed a preference for bounded autonomy, with participants often favoring substantial AI execution while retaining human involvement, oversight, or final authority. We argue that delegation is best understood as a configuration of execution, collaboration, leadership, and authority.
Figures & tables
Figure 1. Overview of the survey procedure and scenario-sampling structure. Participants completed demographic and AI-background measures before evaluating one randomly selected task from each of nine domains, for a total of nine task evaluations per participant. Flow diagram of the survey procedure. Participants first completed demographic and AI-background questions, then evaluated nine task scenarios, with one task randomly selected from each of nine domains. Each participant therefore completed a total of nine task evaluations.
Figure 2. Qualtrics scenario presentation. Participants were shown a task domain and description together with an AI-generated illustration of the activity. Screenshot of the Qualtrics survey interface showing the Home Systems Adjustment scenario presented to participants, including the task description and an AI-generated illustration of the activity. Example Scenario: Home Systems Adjustment .
Code
Definition
Example
%
Assistive / Collaborative Support
Participant asks the AI to help, guide, suggest, explain, analyze, brainstorm, or work alongside them on the task.
“Help me check things and remind me”
66.2
Autonomous Execution
Participant asks the AI to perform the task, take an action on their behalf, or produce the deliverable itself.
“Automatically gauge where I am with minimal input”
48.0
Output / Interaction Specification
Participant specifies the form of the result or how the AI should interact (i.e., using a particular format or channel).
“Scan and analyze, give me a report”
34.4
Reliability, Accuracy & Safety
Participant emphasizes correctness, precision, safety, or trustworthiness of the AI’s work.
“Give me the safest, most optimal route”
17.2
Personalization & Context Awareness
Participant asks the AI to adapt to their preferences, history, constraints, goals, current situation, or data they supply.
“Based on my profile, create a complete plan”
16.4
Proactive / Continuous Assistance
Participant asks the AI to monitor, track, remind, alert, or assist on an ongoing basis rather than in a single exchange.
“Make sure my health is monitored properly”
13.4
Table 1. Codebook for the qualitative content analysis of open-ended responses. Percentages reflect the share of codable responses carrying each code; responses could receive multiple codes, so column percentages do not sum to 100%. A table listing eight qualitative codes with their definitions, a representative example, and the percentage of codable responses to which each code was applied.
Figure 3. Desired AI automation across task domains. Percentage distribution of automation-preference responses within each domain. Percentages sum to 100% within each domain. Chart showing the distribution of participants' desired levels of AI automation across task domains. Each domain is represented by a percentage distribution across five automation-preference categories, ranging from no automation to complete automation. Percentages sum to 100% within each domain.
Figure 4. Trust and attitudes toward AI involvement across domains. Points represent mean ratings and error bars represent 95% confidence intervals. Panel (a) compares trust across agent configurations; Panel (b) summarizes selected considerations relevant to AI delegation. Two-panel figure comparing trust and attitudes toward AI involvement across task domains. Panel (a) shows mean trust ratings for AI agents, human workers, and human-AI teams across domains, with error bars indicating 95% confidence intervals. Panel (b) shows mean ratings for selected attitudes related to privacy, collaboration, and fulfillment across domains, also with 95% confidence intervals.
Figure 5. Willingness to pay for AI services across domains. Distributions are shown separately for one-time and monthly willingness to pay. Because monetary responses were highly right-skewed and included extreme values, the horizontal axis displays log(1+WTP) . Outliers are omitted from the plotted boxplots for readability but were retained in the descriptive analyses. Boxplots showing participants' willingness to pay for AI services across task domains, with separate distributions for one-time payments and monthly payments. The horizontal axis uses the transformed scale log(1+WTP) to accommodate highly right-skewed monetary responses. Outliers are not displayed in the figure but were retained in the descriptive analyses.
Figure 6. Code prevalence by task category. Cells show the percentage of codable open-ended responses within each domain carrying each code. Responses could receive multiple codes; percentages do not sum to 100% within a row. Heatmap showing the prevalence of qualitative codes across task categories. Rows represent task domains and columns represent qualitative codes. Each cell reports the percentage of codable open-ended responses in that domain assigned to the corresponding code, with darker or stronger shading indicating higher prevalence. Because responses could receive multiple codes, percentages within a row may sum to more than 100%.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
#
Block
Items
1
Consent form
1
2
Demographic & AI background
14
3
Task instructions
1
4
Scenarios — Household & Home Management
23
5
Scenarios — Workplace & Professional Support
23
6
Scenarios — Travel
23
Appendix
Table 2. Block structure of the administered survey.
#
Task summary
Description shown to participants
Household & Home Management
1
Recipe Search and Creation
Creating or finding recipes based on preferences and available ingredients.
2
Home Systems Adjustment
Adjusting home lighting, temperature, and energy settings for comfort and cost efficiency.
3
Home Improvement
Planning or figuring out how to complete home improvement or minor repair tasks.
4
Home Safety Monitoring
Checking for signs of potential safety issues at home, such as gas leaks, fires, or intrusions.
5
Entertainment Recommendation
Finding music, movies, or TV content aligned with your tastes.
Appendix
Table 3. Scenario descriptions as presented to participants, domains 1-5 of 9.
#
Task summary
Description shown to participants
Learning & Skill Development
26
Language Learning
Studying and practicing a language by working on vocabulary, grammar, and conversational skills.
27
Studying
Creating flashcards, quizzes, and mind maps from your study materials.
28
Research & Academic Assistant
Reviewing academic papers, summarizing key points, extracting citations, and outlining research topics.
29
DIY/Hands-On Tasks
Finding step-by-step guidance for DIY hobby tasks.
30
Coding
Writing, reviewing, and debugging code, and working through code examples.
Appendix
Table 4. Scenario descriptions as presented to participants, domains 6–9 of 9.
Supplement label
Qualtrics ID
Item
C1
Q1
Consent
D1
Q240
Prolific ID
D2
Q9
Age
D3
Q10
Gender
D4
Q12
Race/ethnicity
D5
Q14
Household income
Appendix
Table 5. Mapping between supplementary item labels and original Qualtrics question IDs.