Intelligent Conversational Agents for Sustainable Tourism Planning: Architecture, Implementation, and Technical Evaluation of an AI Driven Itinerary Generation System

1. Introduction

The global tourism industry stands at a critical juncture, facing the dual challenge of meeting growing demand for personalized travel experiences while addressing urgent environmental sustainability concerns. With tourism contributing approximately 8% to global greenhouse gas emissions and destinations increasingly affected by overtourism [1,2], there is an urgent need for technological solutions that can optimize travel planning while minimizing environmental impact. The World Economic Forum reported that in 2023, the travel and tourism industry showed strong recovery from the COVID 19 crisis, with the Middle East surpassing pre pandemic levels [3], further intensifying the need for sustainable practices.

Recent advances in Artificial Intelligence (AI), particularly in large language models (LLMs) and conversational agents, offer unprecedented opportunities to transform how travelers plan and experience tourism [4,5,6]. These technologies enable natural language interactions that can understand complex user preferences, reason about trade offs, and generate personalized recommendations at scale. The transition from rigid, rule based chatbots to adaptive, agent based frameworks powered by LLMs has revolutionized dialogue systems, enabling them to handle complex multi turn interactions with refined prompting and fine tuning methods [7,8].

However, the integration of AI into sustainable tourism planning remains underdeveloped. Prior research shows that conversational AI is increasingly used in tourism for information retrieval [9], customer support, and itinerary assistance, often improving service efficiency and user experience [5,10,11,12]. Trust, perceived fairness, and emotional resonance also shape adoption, indicating that successful tourism assistants must balance functional performance with user confidence and interaction quality [13,14,15]. In parallel, AI applications for sustainable tourism have focused more often on route optimization, tourist flow management, hotel operations, and environmental monitoring than on user facing itinerary planning [16,17,18,19,20].

Large language models expand this landscape by enabling open ended travel planning, longer horizon interactions, and more personalized recommendation generation [21,22,23]. Yet they remain vulnerable to factual inaccuracies in schedules, availability, and pricing, making tool grounding essential for practical deployment [10,11]. Recent work on tool augmented and agent based LLMs suggests a promising path toward more robust planning systems [24,25,26,27], but validated end to end tourism applications that combine LLM reasoning, external services, and sustainability aware decision making remain scarce [28]. In addition, the environmental footprint of LLM inference itself requires careful assessment when such systems are proposed as sustainability enablers [29,30,31,32].

Against this background, four research gaps motivate the present study: the weak integration of sustainability criteria into conversational travel planning, the limited availability of validated end to end tourism systems, the underexplored problem of reliable tool use under real time constraints, and the lack of empirical understanding of how users respond to sustainability aware recommendations involving trade offs in cost, comfort, and time [13,28].

This paper addresses these critical gaps by presenting a comprehensive study of an intelligent conversational agent specifically designed for sustainable tourism planning. Unlike generic travel chatbots that primarily focus on information retrieval, customer support, and booking oriented assistance [5,10,11], our system actively integrates sustainability criteria into every stage of itinerary generation, from transportation selection to accommodation and activity recommendations. The agent leverages Google’s Gemini 2.5 Flash model, a high performance multimodal LLM with advanced reasoning capabilities and a 1 million token context window [33], combined with real time external service integration through the Amadeus flight API [34]. Amadeus for Developers is a self service travel API ecosystem that provides programmatic access to travel services such as flights, hotels, and itinerary management; in this study, we specifically use its flight resources to retrieve flight offers and related travel data [35].

The system operates at Technology Readiness Level (TRL) 4, representing a validated prototype in a controlled laboratory environment. This development stage is crucial for demonstrating technical feasibility while identifying pathways toward real world deployment. The TRL 4 validation allows for systematic evaluation of core functionalities natural language understanding, context maintenance, external service integration, and sustainability recommendation generation under controlled conditions before advancing to more complex deployment scenarios.

The primary contributions of this research are fourfold:

1.

Novel System Architecture: We present a comprehensive architecture that combines LLMs with external service APIs to create a fully functional conversational agent for sustainable travel planning. The architecture is modular, scalable, and designed for eventual deployment in production environments.

2.

Technical Implementation Details: We provide detailed technical specifications of the system implementation, including prompt engineering strategies, context management approaches, API integration patterns, and sustainability calculation methodologies. These details contribute practical knowledge to the growing field of LLM based agent development.

3.

Systematic Evaluation: We conduct rigorous evaluation of system performance across multiple dimensions: functional correctness, intent recognition accuracy, entity extraction completeness, API integration reliability. Demonstrating that the system meets established targets for TRL 4 validation.

4.

Lessons Learned and Future Directions: We offer practical insights from implementing and testing the system, identifying both technical challenges (prompt consistency, API reliability, carbon footprint estimation accuracy) and opportunities for enhancement (multimodal inputs, real time pricing, booking integration), contributing to the knowledge base for AI driven sustainable tourism applications [36,37].

The remainder of this paper is organized as follows: Section 2 presents the Materials and Methods, including the system architecture, LLM integration, and tool calling workflow. Section 3 reports the experimental results and performance analysis. Section 4 discusses the implications, limitations, and future research directions. Section 5 concludes the paper with the main findings and recommendations for future work.

2. Materials and Methods

2.1. System Architecture Overview

The proposed intelligent conversational agent comprises a conversational frontend interface for user interaction and an LLM centered orchestration layer responsible for natural language understanding, response generation, and tool invocation. The model uses Gemini’s function calling mechanism to decide when to access tools, including the Amadeus flight API and itinerary management functions. This modular architecture, illustrated in Figure 1, facilitates the integration of additional tools while preserving scalability.

The system follows a request response cycle where user inputs in natural language are processed through multiple stages: (1) natural language understanding and intent classification, (2) entity extraction and preference identification, (3) decision making regarding tool invocation and data retrieval, (4) itinerary generation incorporating sustainability criteria, and (5) natural language response generation with structured output formatting. Each stage maintains conversation context, ensuring coherent multi turn interactions essential for complex travel planning tasks.

2.2. Frontend: Conversational Interface

2.2.1. Technology Selection

The conversational interface is implemented using Chainlit, a Python framework specifically designed for building LLM powered conversational applications [38]. Chainlit was selected over alternative frameworks such as Gradio and Streamlit based on several key criteria: native support for streaming responses that enhance perceived responsiveness, built in conversation history management eliminating the need for custom state handling, seamless integration with both LangChain and direct LLM APIs, and professional UI components specifically optimized for chatbot interactions.

The framework provides asynchronous message handling, crucial for maintaining responsive interactions while the language model performs potentially time consuming inference or issues tool calls to external services such as the Amadeus API. The interface supports rich media rendering, though for TRL 4 validation the system focuses primarily on text based interactions to establish core functionality before expanding to multimodal features.

2.2.2. Interface Design Principles

The interface follows conversational design principles aligned with best practices in human computer interaction [39]. Key design decisions include:

Progressive Disclosure: Rather than overwhelming users with comprehensive forms requiring simultaneous input of all travel parameters, the interface reveals information progressively through natural dialogue. The agent asks clarifying questions as needed, gathering destination, dates, number of travelers, transportation preferences, budget constraints, and activity interests through conversational turns.

Context Maintenance: The system maintains complete conversation history within each session, allowing users to reference previous statements without repetition. Users can naturally say "change the activity on the second day to take a walk in the park" or "make it two days shorter" without re specifying which hotel or what duration, as the agent maintains awareness of the current itinerary state.

Error Handling and Recovery: Clear error messages guide users when inputs are ambiguous ("Did you mean Barcelona, Spain or Barcelona, Venezuela?") or when external services are temporarily unavailable ("I’m having trouble accessing flight information right now. Would you like me to continue with general transportation recommendations?"). The system provides actionable recovery paths rather than simply indicating failure.

Feedback Mechanisms: Visual indicators show processing status (typing indicators, progress messages) to maintain user engagement during operations requiring multiple seconds. Confirmation messages summarize understood preferences before generating complete itineraries, allowing users to correct misunderstandings early in the interaction.

2.3. AI Model Integration

2.3.1. Prompt Engineering Methodology

Effective prompt engineering is critical for eliciting desired behaviors from the LLM. Our prompt structure follows a hierarchical design informed by best practices in prompt engineering research [40,41]:

• System Prompt (Role Definition): Establishes the agent’s role, personality, and core behavioral constraints.
• Task Instructions (Dynamic): Specifies current objectives based on conversation state, such as information gathering ("Ask the user about their preferred travel dates"), itinerary generation ("Create a comprehensive 5 day itinerary for Barcelona"), or modification handling ("Update the accommodation to match the user’s budget constraint").
• Context Injection: Includes conversation history (truncated to most recent 50 turns or 100,000 tokens), previously collected user preferences stored in structured format (JSON), and relevant external data (recent flight search results, destination information).
• Output Constraints: Defines format requirements (itineraries in Markdown with day by day information, including schedules and activities), sustainability prioritization rules (prefer trains over flights for sort distances, highlight eco certified accommodations), and content restrictions (avoid promoting environmentally harmful activities).

The prompts evolved iteratively through testing phases, with particular attention to:

• Ensuring a consistent Markdown format for itineraries, so that daily information (schedules, activities) is clear and easy to read.
• Balancing conversational friendliness with efficiency to avoid unnecessarily long exchanges.
• Handling ambiguous or incomplete user inputs gracefully with clarifying questions.
• Prioritizing sustainability without adopting preachy or judgmental tones that might alienate users.
• Maintaining factual accuracy while acknowledging uncertainty when appropriate.

2.3.2. Function Calling Implementation

Gemini’s function calling capabilities enable the agent to invoke external tools when needed. Function definitions follow OpenAPI specifications, allowing Gemini to understand parameters, return types, and appropriate invocation contexts. Implemented functions include:

• search_complete_flights: Queries the Amadeus API for available flights. This integration provides access to real time flight information, including availability, pricing, and estimated $C{O}_2$ emissions. Using actual flight data, the system ensures that users receive accurate and up to date travel options while raising awareness about the environmental impact of each flight. This enables users to make informed and sustainable choices when selecting their itineraries.
• create_itinerary: Generates a complete itinerary based on user preferences and the instructions provided in the system prompt. The function organizes travel elements such as destinations, activities, and dates, ensuring that the final plan aligns with the user’s goals and sustainability considerations.
• modify_itinerary: Updates specific itinerary components based on user requests, maintaining consistency across dependent elements (e.g., adjusting activity schedules when travel dates change).

2.4. Natural Language Processing Pipeline

2.4.1. Intent Classification

The system identifies user intents from natural language inputs to determine appropriate response strategies. Accurate intent recognition is critical, as it ensures that the bot responds appropriately to user needs, avoids misinterpretations, and enables a smooth interaction flow. By correctly identifying the user’s goals, the system can route requests to the right modules.

The main intents detected by the system include:

• InitiateTrip: User wants to start planning a new trip ("I’d like to plan a vacation").
• ProvidePreference: User specifies preferences regarding destination, dates, budget, transportation, or activities.
• RequestItinerary: User asks to generate complete travel plan ("Show me the full itinerary").
• ModifyItinerary: User wants to change specific aspects of generated plan.
• GeneralQuery: User has questions about destinations, sustainability, or travel advice.

2.4.2. Named Entity Recognition

Entity extraction identifies and structures key information from user messages:

• Destinations: Geographic entities including cities, countries, regions (e.g., "Barcelona", "Catalonia", "Spain", "Mediterranean coast").
• Temporal Expressions: Date ranges, relative dates, durations (e.g., "January 5 10", "next weekend", "for five days", "early March").
• Numerical Values: Passenger counts, budget amounts, durations in days.
• Transportation Preferences: Mode specifications (airplane, train, bus, car, bicycle, walking) with qualifiers.
• Activity Interests: Types of desired experiences (museums, hiking, beaches, local cuisine, cultural events).

2.5. Sustainable Recommendation Strategies

The system uses carbon emissions data provided by the Amadeus API for the flights found, accurately reporting emissions for each flight segment. For other itinerary components, such as ground transportation, accommodations, and activities, the model receives specific prompt instructions to prioritize sustainable and eco friendly options.

Transport Prioritization: Modal hierarchy preferring lower emission options when viable: walking/cycling > local trains > regional trains > long distance trains > buses > direct flights > connecting flights > rental cars.

Accommodation Filtering: When suggesting hotels, the model directly prioritizes sustainable and eco certified options. Sustainability features include renewable energy use, water conservation systems, waste reduction programs, local sourcing, and community engagement initiatives.

Activity Recommendations: Emphasis on experiences that minimize environmental impact while maximizing cultural immersion: walking tours led by local guides, visits to conservation projects, nature based activities with educational components, support for local artisans and cultural practitioners, seasonal and regional culinary experiences.

Educational Messaging: Sustainability information presented constructively rather than prescriptively, explaining environmental impacts and benefits without judgment. Users empowered to make informed decisions rather than restricted to predetermined "green" options.

2.6. Evaluation Metrics

To ensure that the performance indicators reported in the Results section are interpretable and reproducible, the evaluation protocol defined a set of metrics before testing the system. These metrics were measured across 45 controlled test scenarios designed to represent common sustainable tourism planning interactions, including new itinerary creation, preference elicitation, itinerary modification, flight search, sustainability comparison, and general destination queries. Each scenario included an expected outcome, a list of required entities, and evaluation criteria for response quality and sustainability integration.

• Functional Correctness (FC): This metric measures whether the agent successfully completed the main task required by each test scenario. A scenario was considered functionally correct when the system produced an itinerary, modification, recommendation, or answer that matched the user’s stated goal and respected the relevant constraints, such as destination, dates, number of travelers, budget, and transport preferences. FC was calculated as the percentage of scenarios completed without major task failure.
• Intent Recognition Accuracy (IRA): IRA evaluates the system’s ability to correctly identify the user’s communicative intention. For each user message, the detected intent was compared with a manually assigned reference label from the intent categories defined in the natural language processing pipeline, including InitiateTrip, ProvidePreference, RequestItinerary, ModifyItinerary, and GeneralQuery. The metric was computed as the proportion of correctly classified intents over the total number of evaluated intent instances.
• Entity Extraction Completeness (EEC): EEC measures how completely the system identified the information required for travel planning from the user’s messages. The evaluation considered destination names, temporal expressions, passenger numbers, budget values, transportation preferences, and activity interests. For each scenario, extracted entities were compared against a manually defined reference set, and completeness was calculated as the percentage of required entities correctly captured by the system.
• Response Coherence (RC): RC assesses the clarity, consistency, and contextual appropriateness of the agent’s natural language responses. Responses were rated on a five point scale, where 1 indicated an incoherent or irrelevant answer and 5 indicated a clear, well structured, contextually consistent response. The final RC value corresponds to the mean score across the evaluated responses.
• API Integration Reliability (AIR): AIR measures the reliability of external service coordination, especially calls to the Amadeus flight API. A tool interaction was counted as successful when the API request was completed, valid data were returned, and the agent incorporated the retrieved information into the response without breaking the dialogue flow. AIR was calculated as the percentage of successful API interactions over the total number of attempted API calls.
• Sustainability Integration (SI): SI evaluates the extent to which sustainability criteria were explicitly incorporated into itinerary generation and recommendation explanations. The assessment considered whether the system included carbon footprint information where available, prioritized lower emission transport alternatives when feasible, suggested eco friendly accommodation or activities, and explained sustainability trade offs in a user understandable way. SI was rated on a five point scale, where 1 indicated no meaningful sustainability consideration and 5 indicated systematic and well explained sustainability integration.
• Average Response Time and Itinerary Generation Time: These latency metrics quantify the system’s responsiveness during interaction. Response time was measured from the moment a user message was submitted until the agent returned a complete answer. Average response time summarizes general conversational responsiveness, while itinerary generation time focuses specifically on complete itinerary construction involving reasoning, tool use, and structured output generation. For both metrics, the reported values correspond to the 95th percentile, reflecting performance under near worst case conditions rather than only average behavior.

3. Results

3.1. System Performance Overview

This section reports the outcomes of the controlled TRL 4 evaluation, focusing on whether the proposed agent could complete representative sustainable tourism planning tasks with the required reliability, conversational quality, and response efficiency. Table 1 summarizes the performance metrics obtained from 45 test scenarios executed over a four week evaluation period.

Overall, the agent met or exceeded all predefined target thresholds, supporting the technical feasibility of the approach at TRL 4. The 92.4% functional correctness rate indicates reliable task completion across diverse scenarios. Natural language understanding was also effective, as reflected in the intent recognition and entity extraction scores, while API integration reliability confirms that external flight data could be coordinated consistently. The response coherence and sustainability integration ratings further suggest that the system combined conversational fluency with domain specific planning requirements.

3.2. Natural Language Understanding Performance

3.2.1. Intent Classification Results

The agent successfully identified user intents in 95.8% of utterances (137 correct classifications out of 143 test cases). Table 2 presents the confusion matrix revealing classification patterns.

Most classification errors occurred in distinguishing between ProvidePreference and GeneralQuery intents (4 cases) when users asked questions while implicitly stating preferences. For example, "What’s the weather like in Barcelona in January?" was occasionally classified as GeneralQuery when the user’s underlying intent was to inform destination and timing preferences. One case confused ProvidePreference with ModifyItinerary when a user refined a previously stated preference.

Examples of successful intent recognition:

Case 1 Complex Preference Statement:

Figure 2. Example of successful intent recognition for a complex preference statement. The screenshot shows a user request specifying the route from Madrid to Barcelona, travel dates from July 5th to 10th, three travelers, a sustainability preference, and a preference for rail transport. The agent interprets this input correctly and summarizes the extracted travel constraints.

Figure 2. Example of successful intent recognition for a complex preference statement. The screenshot shows a user request specifying the route from Madrid to Barcelona, travel dates from July 5th to 10th, three travelers, a sustainability preference, and a preference for rail transport. The agent interprets this input correctly and summarizes the extracted travel constraints.

Case 2 Modification Request:

Figure 3. Example of successful intent recognition for a modification request. The screenshot shows a follow-up message asking to extend the trip until July 12th. The agent correctly identifies the utterance as a modification intent and updates the previously established return date while preserving the conversation context.

Figure 3. Example of successful intent recognition for a modification request. The screenshot shows a follow-up message asking to extend the trip until July 12th. The agent correctly identifies the utterance as a modification intent and updates the previously established return date while preserving the conversation context.

Case 3 Implicit Information Query:

Figure 4. Example of successful intent recognition for an implicit information query. The screenshot shows a user asking wether Barcelona is worth visiting in winter, without explicitly requesting an itinerary. The agent correctly interprets the utterance as an informational intent and provides seasonal destination information while subtly probing for latent travel intent that may support subsequent itinerary planning.

Figure 4. Example of successful intent recognition for an implicit information query. The screenshot shows a user asking wether Barcelona is worth visiting in winter, without explicitly requesting an itinerary. The agent correctly interprets the utterance as an informational intent and provides seasonal destination information while subtly probing for latent travel intent that may support subsequent itinerary planning.

3.2.2. Entity Extraction Performance

Entity extraction achieved 89.2% completeness across all entity categories, with notable variation by entity type as shown in Table 3. Strong performance on destinations (96.7%) and passenger counts (95.6%) reflects clear entity boundaries and limited ambiguity. Date extraction (92.3%) handled various formats successfully ("January 5 10", "first week of January", "5/1 to 10/1"), though occasionally struggled with ambiguous expressions like "mid January" without specific dates.

Budget extraction showed lower accuracy (78.4%) due to varied expression formats:

• Numeric: "1000 euros", "$1500", "1k"
• Qualitative: "budget friendly", "not too expensive", "mid range"
• Contextual: "around what the flights will cost"

The system successfully extracted explicit numeric budgets but sometimes missed qualitative expressions. Future prompt engineering could improve qualitative budget understanding through additional examples.

Activity/interest extraction (86.6%) faced challenges with ambiguous descriptions. "Culture" could encompass museums, historical sites, local traditions, or culinary experiences. The system implemented a strategy of asking clarifying questions when interests were vague:

User: "We’re interested in culture and outdoor activities."

Agent: "Great! When you mention culture, are you more interested in museums and historical sites, or local traditions and food experiences? And for outdoor activities, are you thinking hiking, water sports, or something else?"

3.2.3. Context Maintenance Evaluation

The system maintained coherent context across extended conversations, with test scenarios ranging from 8 to 27 turns (mean: 16.4 turns, SD: 4.8). Context maintenance was evaluated qualitatively across three dimensions:

Anaphora Resolution: Successfully resolved 94.7% (36/38) of pronouns and references:

Turn 8: "What about the hotel options?"

Turn 12: "Can we change it to something cheaper?"

System: Correctly identified "it" as referring to hotel from Turn 8

Preference Consistency: Maintained stated preferences across conversation with 97.8% consistency. In one edge case, when a user first stated "I prefer trains" but later said "We’ll need to fly," the system appropriately updated preferences while acknowledging the change:

Agent: "Understood. I’ll prioritize flight options now. Since you mentioned earlier an interest in sustainable travel, I’ll focus on direct flights which have lower emissions than connections, and I’ll show you the carbon footprint for each option."

Context Switches: Handled topic diversions and returns gracefully in 100% of test cases (8 instances). Example:

Turn 10 (Planning phase): User asks: "What’s the local cuisine like in Barcelona?"

Agent: Provides information about Catalan cuisine

Turn 11: "Thanks! Now back to the itinerary—what about accommodations?"

Agent: Seamlessly returns to itinerary planning without losing state

3.3. External Service Integration Results

3.3.1. Amadeus API Performance

The Amadeus flight API integration demonstrated high reliability with 94.1% successful request completion across 34 flight search operations. Table 4 details API interaction metrics.

The single timeout failure (after 5 seconds) triggered automatic retry logic, succeeding on second attempt with 2.1s latency. The rate limit failure occurred during burst testing when multiple searches were executed simultaneously; the system gracefully fell back to generic flight information and notified the user:

Agent: "I’m experiencing high demand accessing flight information right now. I can provide general flight options between Madrid and Barcelona (typically 1h 15m, prices from 50 150), or we can wait a moment and try again for specific availability."

Table 4. Amadeus API integration performance metrics over 34 flight search operations.

Metric	Value	Notes
Total API Calls	34	Across 12 test scenarios
Successful Responses	32	94.1% success rate
Failed Calls (Timeout)	1	Recovered via retry
Failed Calls (Rate Limit)	1	Handled by fallback
Average Response Time	1.8s	SD: 0.4s
95th Percentile Response Time	2.6s	Within acceptable range
Cache Hit Rate	23.5%	8/34 requests served from cache
CO2 Data Availability	100%	All flights included emissions

Table 4. Amadeus API integration performance metrics over 34 flight search operations.

Metric	Value	Notes
Total API Calls	34	Across 12 test scenarios
Successful Responses	32	94.1% success rate
Failed Calls (Timeout)	1	Recovered via retry
Failed Calls (Rate Limit)	1	Handled by fallback
Average Response Time	1.8s	SD: 0.4s
95th Percentile Response Time	2.6s	Within acceptable range
Cache Hit Rate	23.5%	8/34 requests served from cache
CO2 Data Availability	100%	All flights included emissions

Carbon emission data was available for all successfully retrieved flights, with values ranging from 85kg CO2e (short haul direct flights) to 420kg CO2e (connections with longer routing). The system presented this data prominently:

Agent: "I found 3 direct flight options for your dates:

1. Flight IB3201 Departs 08:30, 89, 92kg CO2e

2. Flight VY2105 Departs 11:15, 76, 88kg CO2e

3. Flight UX7043 Departs 18:45, 94, 95kg CO2e

For comparison, taking the train (3h journey) would emit approximately 12kg CO2e per person, about 87% less than flying."

3.3.2. Simulated Service Performance

Simulated services (accommodation, activities, gastronomy) performed reliably with 100% availability, as expected in controlled environment. Response times were artificially bounded (200 500ms) to simulate realistic network conditions. Data quality validation confirmed that all returned entities included required sustainability attributes (certifications, eco practices, carbon estimates).

3.4. Itinerary Generation Quality

3.4.1. Structure and Completeness

Generated itineraries demonstrated high structural quality across 24 complete itinerary generation tasks. All itineraries included required components:

• Transportation: 100% included with mode, duration, pricing, and CO2 emissions
• Accommodation: 100% included with sustainability certifications and eco practices
• Daily Activities: 100% included with timing, descriptions, and environmental impact notes
• Dining Recommendations: 95.8% included (1 itinerary omitted due to user declining)
• General Sustainability Tips: 100% included at itinerary end

Itinerary length varied appropriately with trip duration: 2 day trips averaged 4 5 activities, 5 day trips averaged 8 12 activities, maintaining realistic pacing without overcrowding schedules.

3.4.2. Chronological Coherence

Temporal sequencing was logically organized in 100% of generated itineraries. The system demonstrated understanding of:

• Meal timing (breakfast, lunch, dinner at appropriate hours)
• Activity duration and transitions (sufficient time for travel between locations)
• Operating hours (museum visits during open times)
• Logical geographic routing (minimizing backtracking)

Example of coherent day structure:

Day 2: Barcelona

09:00 12:00: Sagrada Família guided tour

12:30 14:00: Lunch at nearby organic restaurant Cal Pep

14:30 17:00: Gothic Quarter walking tour

17:30 19:00: Free time for shopping/rest

19:30 21:30: Dinner at farm to table restaurant Granja Petitbo

3.4.3. Personalization Quality

Itineraries reflected stated user preferences with high fidelity. Evaluators rated preference alignment on 5 point scale (1=poor, 5=excellent), achieving mean score of 4.7 (SD: 0.5). Examples:

Budget Adherence: When users specified "budget friendly" travel, the system consistently selected economical options:

• Accommodation: 2 3 star hotels or guesthouses (40 80/night)
• Activities: Free walking tours, public parks, municipal museums
• Dining: Local cafeterias, markets, budget friendly traditional restaurants

Activity Matching: When users expressed interest in "nature and hiking," itineraries emphasized outdoor experiences:

• Mountain excursions (Montserrat, Pyrenees foothills)
• Coastal walks (Mediterranean path sections)
• Parks and gardens (Parc de la Ciutadella, Montjuïc)

Dietary Accommodations: Vegetarian, vegan, and allergy sensitive users received appropriate restaurant recommendations with explicit notes about menu options.

3.5. Sustainable Recommendations Quality

Sustainability integration received mean evaluator rating of 4.8/5.0 across all itineraries. Assessment criteria and results:

Eco friendly Options Prioritized (5.0/5.0): In 100% of cases, when sustainable alternatives existed, they were presented prominently:

• Train recommended over flight for Madrid Barcelona
• Eco certified accommodations listed first with sustainability features highlighted
• Walking/cycling tours prioritized over motorized alternatives

Environmental Trade offs Explained (4.9/5.0): The agent clearly articulated environmental implications of choices in 95.8% of cases:

Agent: "While the direct flight is faster (1h 15m vs 3h by train) and costs about the same, the train journey produces 87% less carbon emissions per person. The train also offers comfortable seating, WiFi, and scenic views of the Spanish countryside."

One case (4.2%) failed to explain trade offs when presenting rental car option alongside public transit.

Local and Seasonal Emphasis (4.7/5.0): Restaurant recommendations emphasized seasonal and local sourcing in 91.7% of cases. Activity suggestions promoted local guides and community based tourism.

Overtourism Awareness (4.5/5.0): When applicable, the agent suggested off peak timing or alternative destinations:

Agent: "Park Güell can be extremely crowded during summer afternoons. I recommend visiting early morning (08:00 09:30) when it’s quieter and you’ll have a better experience while reducing pressure on the site."

3.5.1. Educational Value

Sustainability information was presented educationally rather than prescriptively, achieving positive reception in user feedback simulations. The agent explained environmental concepts without judgment:

Agent: "I notice your itinerary currently includes a flight. Just so you’re aware, this represents about 65% of your total trip emissions. If time allows, the train is a comfortable alternative that would reduce your carbon footprint by about 180kg. However, I understand time constraints are important—what works best for your schedule?"

This approach empowers informed decision making while respecting user autonomy, aligning with best practices in sustainable behavior change [42].

3.6. Performance and Efficiency Metrics

3.6.1. Response Latency

Response times were measured at message level and full itinerary generation level. Table 5 presents latency statistics.

Simple conversational responses (greetings, acknowledgments, clarifying questions) achieved sub 2 second mean latency, maintaining natural conversation flow. Information queries involving knowledge retrieval or light computation averaged 2.3 seconds, within acceptable range for perceived responsiveness.

Flight searches, requiring external API calls and response processing, averaged 3.9 seconds with 95th percentile at 5.1 seconds, just meeting the 5 second target. The maximum observed latency of 6.8 seconds occurred during API retry after initial timeout.

Complete itinerary generation, the most computationally intensive operation, averaged 8.5 seconds with 95th percentile at 12.8 seconds, well within the 15 second target. This includes:

• Multiple LLM inference passes (initial generation, formatting, validation)
• Response structuring and formatting

3.6.2. Token Usage and Cost Analysis

LLM token consumption was tracked across all operations to assess efficiency and project deployment costs. Table 6 summarizes token usage patterns.

Mean conversation cost per complete trip planning interaction (averaging 16 turns including final itinerary generation) was approximately $0.0038, demonstrating economic viability for scaled deployment.

3.7. Error Analysis and System Limitations

3.7.1. Functional Failures

The 7.6% functional failure rate (3 complete failures and 1 partial failure across 45 evaluated scenarios) highlighted several limitations of the proposed system. The most significant issues were related to constraint validation, particularly in situations where user requirements were mutually incompatible, such as high quality accommodation requests combined with very limited budgets. In these cases, the system prioritized generating recommendations over validating feasibility.

Additional limitations were observed in temporal reasoning tasks involving ambiguous natural language expressions. Relative date references occasionally produced incorrect interpretations, especially when requests referred to future months across calendar year boundaries.

The evaluation also revealed a dependency on external service reliability. Failures or delays in the Amadeus API reduced the system’s ability to generate precise flight recommendations, showing that the fallback mechanisms remain insufficient for maintaining consistent functionality under degraded service conditions.

Finally, some generated itineraries did not fully incorporate all user preferences identified during the conversation. This suggests that contextual information extracted during intent recognition was not always consistently propagated through the itinerary generation pipeline.

3.7.2. Entity Extraction Errors

The 10.8% entity extraction error rate (34 missed entities from 363 mentions) broke down as:

• Budget expressions (8 errors): Qualitative descriptions ("moderate budget", "mid range") not mapped to numeric ranges.
• Activity interests (9 errors): Vague terms ("culture", "relaxation") not clarified through follow up questions.
• Temporal expressions (6 errors): Relative dates ("mid month", "around holidays") without specific anchoring.
• Dietary restrictions (4 errors): Implicit restrictions ("I don’t eat meat" not extracted as vegetarian).
• Other (7 errors): Miscellaneous edge cases.

3.7.3. Limitations Identified

Testing revealed several system limitations appropriate for TRL 4:

1.

Real time Pricing Absence: Simulated accommodation and activity pricing does not reflect market rates. Future integration with booking platforms necessary for production deployment.

2.

Availability Checking: System cannot verify actual availability of hotels or activities. Generated itineraries assume availability without confirmation.

3.

Multi destination Complexity: Performance degrades for trips involving 4+ distinct destinations, with reduced personalization quality and increased generation time (mean 18.3s vs. 8.5s for single destination).

4.

Language Limitation: System tested only in Spanish. Multi language support requires additional prompt engineering and potentially fine tuned models.

5.

Visual Content Absence: No integration of images, maps, or visual itinerary representations despite Gemini’s multimodal capabilities. Planned for future TRL phases.

These limitations are documented for addressing in subsequent development phases toward higher TRL levels and eventual production deployment.

4. Discussion

4.1. Key Findings and Implications

This study demonstrates the technical feasibility of intelligent conversational agents for sustainable tourism planning at Technology Readiness Level 4. The system achieved strong performance across key evaluation dimensions, including functional correctness (92.4%), intent recognition (95.8%), and sustainability integration (4.8/5.0). These results suggest that modern large language models, when combined with external service orchestration and domain specific prompting strategies, can support complex multi objective travel planning tasks.

A key contribution of this work is the integration of sustainability considerations throughout the planning pipeline. In contrast to prior tourism oriented conversational systems that primarily emphasize service delivery and personalization rather than explicit sustainability optimization [5,10,11], our approach incorporates environmental impact into multiple decision stages, including transportation, accommodation, and activity selection. The inclusion of carbon footprint estimation provides users with interpretable sustainability information, aligning with prior work emphasizing the role of awareness in promoting sustainable travel behavior [28].

The system also demonstrates reliable interaction with external services, achieving an API integration success rate of 94.1%. This addresses known challenges in tool augmented language model systems related to robustness and error handling. In particular, the use of fallback strategies and retry mechanisms contributes to stable performance under partial service degradation.

In terms of responsiveness, the system achieved an average latency of 3.2 seconds (95th percentile: 5.4 seconds), indicating that real time conversational interaction is feasible even when combining language model inference with multiple external API calls. Techniques such as asynchronous execution and caching were important in maintaining acceptable response times for user interaction.

4.2. Practical Implications for Industry

The findings have several implications for stakeholders in the tourism sector:

Travel Agencies and Tour Operators: The proposed architecture illustrates how conversational AI can automate routine itinerary generation, allowing human agents to focus on complex or high value interactions. The reported cost per conversation suggests potential economic feasibility for large scale deployment.

Destination Management Organizations: The presentation of sustainability information in an educational and non restrictive manner provides a mechanism for encouraging more environmentally conscious decisions without reducing user autonomy.

Policy Makers: The integration of carbon estimation into travel planning systems highlights the potential of AI tools to support evidence based environmental policy in tourism contexts [19].

4.3. Environmental Impact Considerations

While the system is designed to promote sustainable travel behavior, its own computational footprint must be considered. Recent literature indicates that the environmental cost of LLM use depends strongly on the inference setup, hardware, prompt design, and workload characteristics; therefore, generalized per conversation CO₂e estimates should be treated with caution [29,30,31].

To contextualize this impact, it remains significantly lower than emissions associated with common travel choices such as short haul flights. However, the net environmental benefit depends on whether the system effectively influences user decisions toward lower emission alternatives. As such, conclusions about overall sustainability impact should be interpreted cautiously and require further empirical validation at scale.

Ongoing improvements in data center efficiency and increased use of renewable energy are expected to reduce the environmental cost of AI systems over time [19].

4.4. Future Research Directions

Several directions emerge for future work:

• Real world validation: Deployment studies with actual users (TRL 5 6) to evaluate behavioral impact and system usability in realistic settings.
• End to end booking integration: Extension of the system to support full transaction flows, including reservation, payment, and confirmation handling.
• Multimodal interaction: Incorporation of visual and spatial information to improve travel planning experience and engagement.
• Personalization mechanisms: Development of long term user modeling approaches while ensuring privacy preserving data handling.
• Overtourism Management: Developing predictive models that assess destination capacity and recommend alternative timing or locations when popular sites experience pressure, contributing to more equitable tourism distribution [17].

5. Conclusion

This paper presented a TRL 4 intelligent conversational agent for sustainable tourism planning that combines large language model capabilities with external tool use to generate personalized itineraries informed by sustainability criteria. The proposed system shows that LLM based conversational planning can move beyond general travel assistance toward structured decision support that integrates transportation options, accommodation suggestions, activities, and carbon awareness within a single interaction flow.

The technical evaluation indicates that this approach is feasible under controlled conditions. Strong results in functional correctness, intent recognition, response quality, and API reliability suggest that an LLM centered architecture can sustain multi turn travel planning while grounding critical information in external services when needed. These findings are especially relevant for tourism applications, where user requests are inherently open ended but operational details such as flight options require dynamic data access.

Beyond technical feasibility, the study contributes an application oriented perspective on sustainable tourism AI. Rather than treating sustainability as a separate reporting layer, the system embeds it directly into itinerary generation and recommendation logic. In this sense, the work illustrates how conversational agents can simultaneously support hyper personalization and encourage more environmentally responsible choices without removing user autonomy from the planning process.

At the same time, these conclusions must be interpreted in light of the current prototype scope. The controlled TRL 4 setting, dependence on external services, simulated data for some itinerary components, and absence of real world user deployment limit the generalizability of the results. Future research should therefore examine behavioral impact, multimodal interaction, broader tool ecosystems, and operational deployment with real travelers and stakeholders.

Overall, this study provides evidence that tool augmented LLM systems can serve as a credible foundation for next generation travel assistants that are both personalized and sustainability aware. By bringing together conversational AI, external travel data, and environmental considerations in a single architecture, this work establishes a concrete basis for future research and development toward practical intelligent tourism planning systems.