Source-linked AI summary
How Constraints and Preferences Shape Travel Planning: Implications for AI Planning Support
Fuling Sun, Yining Cao, Peiling Jiang, Mingyi Li, Haijun Xia
TL;DR
AI planning tools often assume users can state stable goals upfront, although planning needs may evolve as people explore. Through interviews with travelers and travel agents, the paper characterizes these dynamics and identifies recurring actions and design implications for collaborative planning support.
Problem
AI planning tools often assume users can articulate stable goals upfront, while real-world planning is iterative and constraints and preferences evolve.
Source: §1 · p. 1Method
The study interviewed eight travelers about planning and AI tools and nine travel agents about their professional planning practices.
Source: §1 · p. 1Results
The authors characterize three types of constraints and preferences and identify 11 recurring actions, alongside gaps in how current AI systems support travel planning.
Source: §7 · p. 14Takeaways & Limitations
AI planning systems should represent constraints and preferences as evolving elements and structure collaboration to support their discovery, coordination, and revision.
Source: §1 · p. 2, §7 · p. 14Convenience sampling and brief AI-tool interactions yielded snapshots rather than complete end-to-end planning processes, limiting representation across populations and real-world planning duration.
Source: §3.4 · p. 5
Abstract
from arXiv · showhide
Planning is a common yet complex activity shaped by constraints to satisfy and preferences to balance. Travel planning, as both an everyday activity and a frequent benchmark for evaluating intelligent systems, offers a rich context for examining how constraints and preferences emerge and evolve. While recent AI systems have achieved impressive results in generating personalized itineraries, they often assume that users can articulate stable goals upfront. To understand how real-world planning unfolds, we conducted a two-part interview study: one with eight travelers reflecting on their planning experiences, and one with nine travel agents sharing professional practices. We trace the dynamics of constraints and preferences as they are surfaced, refined, and coordinated throughout the planning process, and identified 11 actions revolving around constraints and preferences, which shaped the planning process. We offer design heuristics for planning tools that better support human-AI collaborative actions to support the fluid, contingent nature of planning.
1 Introduction
Planning is iterative, shaped by constraints and situated judgments, yet AI planning tools often assume users can state stable goals upfront. This study examines how constraints and preferences evolve in travel planning and what that means for human-AI planning support.
1 Introduction
- The study asks how constraints and preferences shape travel planning and what these dynamics imply for future AI-powered planning systems.It interviews eight travelers about recent planning and AI tools, then interviews travel agents about their practices.Source: §1 · p. 1
- The authors identify 11 actions around constraints and preferences, reflecting travelers’ needs, agents’ assistance, and gaps in current AI-assisted planning.The study traces how these elements are surfaced, refined, and coordinated throughout planning.Source: §1 · pp. 1–2
- Travelers’ needs and preferences emerge through exploration, while AI tools often rush to generic itineraries and rarely clarify or resolve conflicts.Travel agents instead probe preferences, identify ambiguities, and negotiate to compose feasible, personalized plans.Source: §1 · p. 2
- Planning tools should model constraints and preferences as evolving, visible, manipulable representations and support collaborative articulation, refinement, and reconciliation.Users should be able to revisit inputs, adjust priorities, and explore alternatives as understanding develops.Source: §1 · p. 2
2 Related Work
Planning research spans formal goal-directed approaches and situated accounts of adaptive action, while travel planning offers a setting where decisions and needs evolve. Work on AI planning and personalization has begun supporting user input, but coordination among changing, potentially conflicting needs remains under-explored.
2.1 Understanding Planning and Travel Planning
- Everyday planning is dynamic and iterative, while formal automated planning typically assumes known constraints represented in a predefined model.Situated-action accounts instead emphasize adaptation to local contingencies; travel planning involves interdependent decisions and changing conditions.Source: §2.1 · p. 2
2.2 Planning with Large Language Models
- LLM planning research explores human feedback and plan editing, but coordination of multiple, potentially conflicting needs remains under-explored.Travel planning makes this interplay salient because diverse constraints and preferences must be managed in context.Source: §2.2 · p. 3
2.3 Constraints, Preferences and Personalization
- Users’ constraints and preferences can be under-specified, emerge or shift, and interact, so personalization must address their evolving relationships.Rather than seeking one correct answer, people adaptively balance needs and negotiate trade-offs; the paper aims to make these dynamics visible and actionable.Source: §2.3 · p. 3
3 Methods
The study draws on two perspectives to understand travel planning: travelers planning their own trips and professional agents planning for clients.
3 Methods
- The researchers sought insights from experienced travelers who plan their own trips and professional travel agents who plan for clients.This paired perspective was intended to capture end-user experiences and expert assistance strategies.Source: §3 · p. 3
- The traveler perspective focused on people’s own travel-planning experiences.It was paired with the professional-agent perspective within the same travel-planning domain.Source: §3 · p. 3
- The agent perspective focused on professional strategies for assisting clients with travel planning.Together, the two perspectives capture both end-user practices and expert assistance.Source: §3 · p. 3
3.1 Perspective 1: Learning from Travelers
The traveler study combined interviews about recent trips with hands-on interactions with two AI-assisted travel-planning tools. The sessions elicited reflections on how tools supported or constrained planning, rather than evaluating completed itineraries.
3.1.1 Participants.
- Eight frequent travelers with recent and upcoming trips were recruited through convenience sampling to discuss their planning practices.Participants had prior experience with AI tools but had not used AI to plan previous trips; the study was not designed to assess prompting skill or expert AI use.Source: §3.1.1 · pp. 3–4, §3.1 · p. 3
3.1.2 AI System Selection.
- Gemini Pro represented conversational planning and Tripadvisor represented a domain-specific graphical interface; the tools prompted reflection, not a performance comparison.Participants tried both tools, one at a time, for upcoming trips.Source: §3.1.2 · p. 4
3.1.3 Procedure.
- Participants described recent planning and then used two AI tools while thinking aloud, with tool order randomized to mitigate order effects.Each interaction lasted about 30 minutes, shorter than real-world planning that may unfold over days or weeks; participants were not expected to complete an itinerary.Source: §3.1.3 · p. 4
- The interviews sought to examine how travelers articulated, negotiated, and revised constraints and preferences during AI-assisted planning, not the quality of the resulting plans.Each interview lasted about 90 minutes and was conducted over Zoom with screen sharing.Source: §3.1.3 · p. 4
3.2 Perspective 2: Learning from Travel Agents
The study examined expert-mediated travel planning through interviews with nine travel agents, focusing on how they work with clients to create customized itineraries.
3.2 Perspective 2: Learning from Travel Agents
- The study framed expert-mediated planning around how agents address constraints and preferences.Figure 1 contrasts Tripadvisor’s structured destination-to-itinerary workflow with Gemini’s varied question-and-planning requests.Source: §3.2 · p. 4, Figure 1 · p. 5
3.2.1 Participants.
- Nine agents participated: two specialized in corporate travel, while seven focused on leisure travel.Participants were recruited from an industry catalog and partner agencies based on positive reviews and active professional practice.Source: §3.2.1 · p. 4
3.2.2 Procedure.
- Semi-structured interviews explored agents’ recent client plans, communication, itinerary customization, and reflections on their roles.Interviewers asked about similar situations beyond the initial example and invited agents to share relevant planning materials.Source: §3.2.2 · p. 4
- Interviews lasted 60–90 minutes and were conducted in person with two agents and remotely via Zoom with seven.Agents received an $80 Amazon gift card as compensation.Source: §3.2.2 · pp. 4–5
3.3 Analysis
The researchers analyzed traveler and agent interviews separately before synthesizing themes across the two participant groups.
3.3 Analysis
- Iterative coding surfaced patterns in how constraints and preferences were expressed and negotiated.Codes remained provisional and were revised as interpretations developed through research-team discussions.Source: §3.3 · p. 5
- The team compared patterns across participant groups and refined overarching themes to capture shared dynamics and contrasting perspectives.The first author organized candidate themes, which the team reviewed and refined before cross-group synthesis.Source: §3.3 · p. 5
3.4 Limitations
The study offers an empirical basis for examining constraints and preferences in travel planning, while its sampling and brief tool interactions limit the scope of its findings.
3.4 Limitations
- Convenience sampling may not represent travel planning practices across cultures, socioeconomic backgrounds, and travel styles.The authors identify these groups as areas where participant diversity may be limited.Source: §3.4 · p. 5
- Brief interview-time interactions with AI tools captured snapshots rather than complete end-to-end planning processes.Real-world planning often unfolds over days or weeks, which the study interactions did not fully capture.Source: §3.4 · p. 5
- Despite these limits, the study provides an empirical basis for examining travel-planning constraints and preferences and informing AI-supported tools.Source: §3.4 · p. 6
4 Findings: The Dynamics of Constraints and Preferences in Travel Planning
This section examines how constraints and preferences arise from external conditions, travelers’ priorities, and relationships among plan components, and evolve through iterative planning actions rather than remaining fixed or following a linear process. It also considers how these dynamics relate to AI tools’ difficulties with conflicts, stated interests, and changing criteria or circumstances.
4.1.1 External Factors: Destination Constraints and Offerings.
- Destination availability, schedules, and travel logistics shape what is feasible, and travelers may need to adjust plans when they overlook these external constraints.Travel agents described clients underestimating travel distances and destination conditions, while one traveler changed dates to catch whale-watching season.Source: §4.1.1 · p. 6
4.1.2 Internal Factors: Travelers’ Constraints, Preferences, and Inten-
- Personal constraints, preferences, and goals shape both plan feasibility and whether an experience feels meaningful, with similar circumstances producing different priorities across travelers.For repeat visits, one traveler wanted to revisit places tied to memories, while another preferred to avoid previously explored areas.Source: §4.1.2 · p. 6
4.1.3 Integrative Factors: Coherence Across the Plan.
- Integrative factors connect activities and create dependencies: fixed bookings constrain later choices, while pacing, variety, and balance affect whether an itinerary is coherent.An AI itinerary with four museum visits in three days was criticized despite the traveler’s general interest in museums.Source: §4.1.3 · p. 6
4.1.4 Interplay among Factors.
- The three factor types are flexible and overlapping, and a consideration may shift across them without following a clear or linear sequence.A sunset goal began as a personal intention, became tied to sunset timing, and then constrained how other activities were arranged.Source: §4.1.4 · p. 6
4.2 Actions Revolving around Constraints and Preferences
- Travelers and agents surface preferences through direct specification, examples, and past-experience narratives, but travelers may struggle to articulate vague or evolving needs upfront.Agents use intake forms and conversations about memorable trips; travelers inspect concrete itineraries and shared experiences to recognize what they want. AI summaries can reduce research effort but narrow early exploration.Source: §4.2.1 · p. 8, §4.2.2 · p. 8, §4.2.3 · p. 8
- Refining and visualizing preferences turns broad needs into actionable criteria and inspectable plans, though AI categories can misinterpret intent and uniform itineraries can obscure priorities.Travelers and agents use situational evidence and follow-up questions to clarify requirements, while structured schedules expose timing and commitments; participants found AI-generated step-by-step lists hard to inspect for priorities.Source: §4.2.4 · pp. 8–9, §4.2.5 · p. 9
- Anchoring high-priority or time-sensitive requirements helps order plans, while referencing why options were chosen supports interpretation and informed revision.Travelers and agents prioritize key events and limited availability, but tested AI systems sometimes failed to protect an unstated conference schedule; travelers also questioned recommendations without clear justification.Source: §4.2.6 · p. 9, §4.2.7 · pp. 9–10
- Evaluating individual requirements and detecting conflicts helps identify unrealistic or inconsistent requests, but AI tools often accepted impractical inputs without questioning them.Agents questioned infeasible requests and proposed alternatives, whereas AI-generated plans could leave travelers to discover feasibility problems after generation; the passage introduces conflict detection as a further action.Source: §4.2.8 · p. 10, §4.2.9 · p. 10
- Gemini often missed conflicts between requests, while AI tools also overfit expressed preferences into narrow, unbalanced itineraries.For example, Gemini included a southern waterfall despite P4’s stated focus on the north; preference-heavy plans left little room for variety or adjustment.Source: §4.2.9 · p. 10
- Curation used constraints and preferences to narrow options, but travelers’ emerging interests and AI tools’ fixed initial keywords could make selection difficult to adjust.Travelers used emerging preferences to guide discovery, while agents offered tailored, limited alternatives; Tripadvisor’s keyword-based workflow could require restarting when interests changed.Source: §4.2.10 · p. 10
- Travelers left decisions open for situational changes, and agents prepared alternatives, whereas AI plans could be fixed and offered limited support for uncertainty.The study did not observe real-world execution of AI-generated plans; travelers described deferring restaurant choices to adapt to on-site conditions, and agents used alternatives to accommodate individual and group needs.Source: §4.2.11 · pp. 10–11
4.2.12 Interplay among Actions.
- Planning actions were interwoven and iterative, so failures in conflict detection or decision justification could propagate and require travelers to search for options themselves.Visualizing plans could enable anchoring priorities and detecting conflicts, while discovering constraints co-occurred with curating options.Source: §4.2.12 · p. 11
5 Design Implications: Designing for Evolving Constraints and Preferences in Planning
The study frames constraints and preferences as evolving elements of planning and proposes design implications for modeling them over time and supporting collaborative action. Together, these implications treat planning as an iterative process of surfacing, refining, and negotiating requirements.
5.1 Modeling Evolving Constraints and Preferences
- The authors propose modeling constraints and preferences as evolving throughout planning, with three design considerations for representing and supporting that evolution.Their analysis treats these elements as continuously surfaced, refined, and applied rather than fixed inputs at the outset.Source: §5.1 · p. 11, §5 · p. 11
5.1.1 Presenting Persistent Constraints and Preferences.
- Planning systems should model constraints and preferences as evolving objects that remain visible and adjustable throughout the process.Travelers could revisit earlier inputs, relax or strengthen constraints, and reorganize priorities as their understanding changes.Source: §5.1.1 · p. 11
5.1.2 Capturing Constraints and Preferences in Context.
- Systems should capture emerging needs in context by supporting browsing, interaction-based annotation, and confirmation or refinement of candidate constraints.For example, interaction cues such as highlighting cost, dates, or activity types could suggest annotations without interrupting exploration.Source: §5.1.2 · p. 11
5.1.3 Configuring Interactive Constraints and Preferences.
- Adjustable constraint tokens with priority and flexibility attributes could let travelers test trade-offs and see how changes affect itinerary feasibility.Systems could also propose alternative interpretations of abstract preferences for travelers to select or refine.Source: §5.1.3 · p. 11
5.2 Supporting Collaborative Actions on Constraints and Preferences
- Several planning actions remain unsupported or only partly handled by current AI tools, limiting users’ ability to clarify needs, resolve conflicts, and steer evolving preferences.The authors recast these actions as a vocabulary for structuring human–AI co-planning rather than assigning them solely to users or AI.Source: §5.2 · p. 12
- Collaborative actions can unfold in sequence: evaluating a plan may reveal violations that prompt requirement refinement, while discovery may lead to reprioritizing and regenerating options.The authors argue that supporting these transitions can distribute initiative and responsibility between users and AI.Source: §5.2 · p. 12
5.3 Summary
- The design implications treat constraints and preferences as shared objects of work and offer a transferable lens for complex planning beyond travel.The approach emphasizes both representing evolving requirements and explicitly supporting actions that surface, refine, and negotiate them.Source: §5 · p. 11, §5.3 · p. 12
6 Discussion
The discussion argues that the observed planning dynamics may extend beyond travel to ill-structured tasks with open-ended goals and evolving constraints. It also identifies boundaries in the study’s participant sample and exploratory use of existing AI tools.
6.1 Beyond Travel Planning
- The authors expect the findings to generalize to ill-structured planning with open-ended goals, evolving constraints, iterative revision, and uncertain environments—not fixed, fully specified tasks.They frame planning systems as participants in ongoing construction of the planning space and describe collaborative actions as a domain-agnostic lens.Source: §6.1 · p. 12
6.2 Understanding Human-AI Practices
- Comparing travelers, professional agents, and current AI systems within one domain reveals which actions AI supports and where collaboration breaks down.The task-level framing examines how responsibility, initiative, and interpretation are negotiated in practice.Source: §6.2 · p. 12
6.3.1 Participant Sample.
- The convenience sample came from the authors’ accessible networks and may not represent infrequent travelers or people from different cultural and socioeconomic backgrounds.The authors suggest broader, cross-cultural, or large-scale studies to examine varied planning contexts.Source: §6.3.1 · p. 12
6.3.2 Study Methodology.
- The study used less-structured exploration of existing AI tools, so its interaction patterns may not anticipate practices that emerge with greater familiarity or deeper workflow integration.The authors suggest longitudinal studies or in-situ deployments to observe those developments.Source: §6.3.2 · pp. 12, 14
7 Conclusion
The conclusion reports that constraints and preferences evolve throughout travel planning, while current AI tools reveal gaps in supporting that process. It proposes modeling evolving requirements and structuring human–AI collaboration around actions on them.
7 Conclusion
- The study identifies three types of constraints and preferences and 11 recurring actions, while finding gaps in current AI support for travel planning.The authors characterize these elements as identified, refined, and applied throughout planning rather than fixed at the outset.Source: §7 · p. 14
- The authors propose design implications for modeling evolving constraints and preferences and a vocabulary of collaborative actions for structuring user–AI interaction.The proposed framing treats requirements as first-class objects for mixed-initiative discovery, coordination, and revision.Source: §7 · p. 14
- The authors present broader lessons for AI systems supporting complex planning tasks with open-ended goals and evolving requirements, while grounding the findings in travel planning.Source: §7 · p. 14