[Case 01]
TravelGPT: Designing an AI Workflow for Human Collaboration
TravelGPT: Designing an AI Workflow for Human Collaboration
TravelGPT: Designing an AI Workflow for Human Collaboration
AT-A-GLANCE
TravelGPT (formerly VOYA) is an AI-powered itinerary builder that reimagines how travel agents plan and personalize trips. The mission was to help agents go from hours of manual planning to an intelligent, collaborative experience where they could co-create itineraries with AI.
The challenge: designing for a technology (agentic AI) that was still being defined; while grounding it in real-world UX principles and trust.
[Industry]
Travel Planning
[My Role]
Lead UX Designer
[Platforms]
Desktop
[Timeline]
October 2024 - Present
Objectives
Redefine itinerary creation as a collaborative AI-human workflow.
Build an interface that bridges system reasoning with human decision-making.
Reduce manual effort and improve planning accuracy, transparency, and speed.
Key Outcomes
90% reduction in itinerary creation time (5 hours → under 5 minutes).
78% of beta users described the experience as “magical.”
Served as a foundation for TBO’s AI strategy and luxury B2C travel roadmap.
Who are we solving for?
Travel agents are professional intermediaries who design, plan, and manage trips for individuals, families, and organizations. They combine market knowledge, supplier networks, and human judgment to curate travel experiences that match clients’ needs and budgets. In an era dominated by online booking, travel agents remain essential because they provide personalized guidance, handle complex itineraries, and offer trusted post‑booking support. TBO’s user base includes these professionals globally — from independent consultants to enterprise agencies — serving leisure travelers, corporate clients, and luxury seekers alike.
Designing for TravelGPT meant building for a diverse, international ecosystem of travel experts, each with distinct workflows, cultural contexts, and technological maturity levels.
The current state of travel planning
In most agencies today, travel planning remains a heavily manual process. A typical travel professional works across spreadsheets, chat threads, and vendor portals—coordinating flights, hotels, transfers, and activities while juggling client preferences and supplier updates. Information is fragmented, and each new request requires a mix of copy‑paste operations, negotiation calls, and visual formatting. There are no dedicated tools that unify these steps or preserve context. Every itinerary becomes a small project in itself, reliant on the agent’s personal organization, intuition, and recall.
This fragmented workflow not only slows productivity but also leaves little time for what matters most: curating experiences that feel personal and memorable.
Background and Problem Definition
The Beginning
It all started with a vision: What if AI planned travel?
Since we were diving into uncharted waters of new technology within the travel planning domain, I organized a small design hackathon where the UX and product teams were divided into smaller groups. The goal was to think boldly, explore freely, and bring unconventional ideas to the table that tested the limits of what AI could do for travel.
This exercise helped us gather a wide range of concepts some practical, some wildly futuristic all of which allowed us to map overlaps and identify recurring pain points. From these, we distilled common problem statements that could define the foundation of the project moving forward.
Early Validation
To validate the common problem statements identified during the design hackathon, I interviewed 10 travel agents across markets (India, UAE, UK), focusing on.
How they currently build itineraries
Which parts of their workflow they found repetitive
Where they felt “creative satisfaction” versus “manual frustration”
Competitor Benchmarking
I analyzed products like TravelPerk, Travefy, and TripHobo, studying their flows, UX patterns, and gaps:
Most tools were linear and rigid, optimized for travelers, not travel agents.
None allowed AI-powered collaboration or real-time itinerary structuring.
This uncovered our opportunity space — AI as an orchestrator, not just a formatter.
Identified Problem Statements
Ideation & Early Wireframes
Armed with the insights, I brought the team together once again to share the research findings and key observations. We conducted a collaborative brainstorming session aimed at exploring a wide spectrum of design directions, encouraging everyone to sketch, critique, and iterate rapidly. The goal was to produce multiple ideas and perspectives on how the product could evolve. Some examples are:
Text-to-itinerary via prompt input
Visual card-based drag-and-drop assembly
AI auto-layout based on trip length and destinations
After capturing and synthesizing all ideas, I distilled the strongest concepts and moved forward into the design phase with clear, validated directions.
Outcome
I built the first version of VOYA, a 100% Figma prototype where I simulated the end-to-end user experience. Some of its core interface elements were converted into a working prototype using Lovable and Supabase, allowing us to test real interactions.
VOYA v1 was then taken into the field, where I spent two weeks in Mumbai working alongside travel agents in their offices — observing, testing, and participating in their real workflows.
This immersion helped uncover deep use cases, revealed gaps between design and daily operations, and captured authentic first reactions to how AI could fit into their existing routines.
Quick Summary:
Highlights
Began with a bold vision of “AI-planned travel” launched through an internal TBO design hackathon.
Conducted 10 validation interviews with agents, followed by multiple brainstorming rounds to refine problem statements.
Benchmarked competitors (Travefy, TripHobo, TravelPerk) to uncover unmet gaps.
Solution
Facilitated cross-team ideation sessions to explore diverse possibilities and identify recurring pain points.
Led design sprints that produced rough sketches, wireframes, and multiple concept iterations tested with real agents.
Built VOYA v1 — initially a Figma prototype with select components prototyped using Lovable and Supabase — and field-tested it for two weeks in Mumbai with travel agents.
Results
Established a shared understanding of travel workflow challenges across teams.
Captured authentic early feedback that guided the first working prototype — VOYA v1.
60% faster output, while field testing uncovered deeper challenges in the workflow that shaped the next phase.
Research and Insight
The two-week immersion in Mumbai proved extremely useful. Users were visibly excited about the prospect of AI in travel planning, yet they felt the solution only scratched the surface of their real workflow challenges.
I gathered all the positive feedback and pain points from the Mumbai agencies and decided to expand the research globally.
Research Plan
30 structured interviews with agents from multiple regions. This time, the interviews would go deeper—focusing on specific, critical segments of the travel agent workflow to uncover opportunities for more meaningful problem solving.
We needed to understand why agents weren’t ready to adopt it deeply and what “value” truly meant to them.
Processing the Insights
We transcribed conversations, coded responses, and built a Jobs-To-Be-Done (JTBD) journey map to visualize what agents were trying to get done, not just what they were doing.
Through this lens, it became clear:
Agents didn’t need AI to make things faster; they needed AI to remove friction from multi-step coordination.
They valued control and confidence over automation.
Refined Problem Statements: The Depth of AI Opportunity
Building on insights from the global interviews and field immersion, three new problem statements emerged that went beyond the earlier “beautification” stage. These reflected the structural and systemic challenges agents faced in their day-to-day workflows.
These redefined problem statements elevated our design challenge from simply improving interface efficiency to architecting an AI system that could orchestrate, explain, and personalize travel planning holistically.
We also realized that the initial problem statements only addressed a narrow portion of the agent’s workflow — mainly the beautification phase. Through the JTBD journey map, it became clear that the deeper issues stretched across the entire lifecycle of travel planning, from understanding the traveler to final delivery. This realization shifted our design ambition from optimizing a single step to orchestrating an ecosystem-wide workflow solution.
These findings reshaped our north star: build an AI that collaborates, not replaces.
Quick Summary:
Highlights
Conducted 30 remote interviews and 2 immersive field studies, starting with two-week Mumbai immersion and expanding globally.
Synthesized insights through affinity mapping and qualitative coding.
Solution
Used JTBD mapping to reveal how initial problem statements only solved part of the agent’s journey and redefined the scope to span the full workflow.
Derived three new global problem statements around fragmented decision-making, opaque AI logic, and scalability of personalization.
Results
Highlighted need for an AI system that orchestrates the entire planning lifecycle, not just beautification.
Anchored future design phases in validated behavioral and strategic data from real agents.
Redefining the Product
Challenge
The core insight reframed VOYA’s purpose — it was no longer a beautifier but an intelligent workflow assistant. Around this time, “Agentic AI” was emerging as a design paradigm: systems composed of multiple micro-agents working together, much like human teams.
Agentic Architecture
Drawing from this, we defined a multi-agent orchestration model, where each AI unit specialized in a domain:
Routing Agent: Suggested optimized city orders and travel modes.
Hotel Agent: Recommended stays based on availability and preferences.
Activity Agent: Curated day-wise experiences.
Optimizer: Balanced pacing, transitions, and overall cohesion.
AI UX Challenge
Designing for parallel AI reasoning required creating explainable UX layers:
Introduced visual reasoning overlays to show how AI made choices.
Added user intervention points for human override.
Built a feedback loop to “teach” AI through real user corrections.
We also defined AI personality boundaries, balancing conversational tone with operational precision. The assistant needed to feel intelligent, but not intrusive.
Outcome
VOYA v2 (now internally called TravelGPT) became an agentic workflow platform, capable of not only generating itineraries but optimizing and contextualizing them live.
Quick Summary:
Highlights
Reframed product as a multi-agent workflow system.
Defined AI-human collaboration principles.
Solution
Designed an Agentic AI UX — transparency, reasoning trace, and user control.
Results
Built scalable agent framework across routing, hotels, and activities.
Design and Development
Design Approach
We adopted a build–test–refine loop, working in tight collaboration between design and engineering. The team consisted of two developers and myself, operating without PMs aligning daily through prototypes and functional experiments.
Core UX Concepts
Day-Card Layout: allowed modular drag-drop planning and parallel AI updates.
Smart Routing: users could accept, reject, or edit AI-generated sequences.
Explainable Prompts: conversational context displayed decision rationale.
Real-Time Collaboration: live agent–client sharing and co-editing.
Designing for AI UX
A key challenge was designing for non-linear intelligence. Unlike traditional systems that progress step-by-step, AI can think in parallel. To accommodate this, the interface needed to:
Handle multiple evolving suggestions at once.
Indicate which AI agent was “thinking.”
Maintain user trust by keeping outputs predictable, not magical.
We created anticipatory states (loading cues + transparent micro-feedback) to signal progress without breaking flow.
Beta Rollout
The beta version launched with travel agents across India, APAC, and UAE.
78% found the experience intuitive.
60% adopted it as their default workflow.
90% reduction in overall planning effort.
Quick Summary: Research Phase
Highlights
Operated as a lean 3-member team.
Integrated parallel AI reasoning into interface.
Solution
Built modular Day-Card UX and explainable AI prompts.
Designed feedback loops and real-time collaboration tools.
Results
90% faster planning, 60% workflow adoption.
78% user satisfaction in closed beta.
Market Adaptation
The Problem
When TravelGPT scaled to less tech-forward markets, adoption dipped. Agents expressed hesitation, fearing AI might replace their expertise.
Design Response
We re-examined the onboarding experience and reframed the narrative:
From “AI builds for you” → “AI builds with you.”
Added explainer steps, editable checkpoints, and real-time AI transparency.
Reduced cognitive friction with micro-interactions showing what the AI was doing.
Impact
+25% trust improvement in post-onboarding surveys.
+40% increase in repeat usage.
A smoother learning curve for first-time AI users[Image: Amplitude Funnel]
Quick Summary:
Highlights
Noticed adoption drop in non-tech markets.
Identified that perception, not performance, was limiting adoption.
Solution
Reframed UX around human-in-the-loop AI and transparency.
Redesigned onboarding and explainability cues.
Results
25% trust and 40% repeat usage improvement.
Boosted cross-market confidence and retention.
Strategic Impact
Beyond Product
TravelGPT’s architecture proved adaptable beyond travel agents. Its agentic foundation could support B2C personalized travel planning, positioning it as TBO’s competitive moat for future expansion.
Design as Strategy
I developed conceptual models for how AI could scale to serve both travelers and agents, building ecosystem cohesion through shared intelligence.
Strategic Outcome
Design matured from execution to strategy — influencing product vision, roadmap prioritization, and even how leadership imagined TBO’s AI-powered future.Handle multiple evolving suggestions at once.
Quick Summary: Research Phase
Highlights
Identified platform potential for B2C evolution.
Elevated design’s role in shaping business vision.
Solution
Proposed “World of TravelGPT” connecting agents, travelers, and AI.
Created scalable interaction architecture.
Results
Influenced TBO’s luxury travel roadmap.
Reframed design as a growth catalyst.
Key Learnings
Beyond Product
Principle-first design sustains technology-first products. Even when AI logic evolved, UX remained rooted in human clarity, trust, and control.
Empathy scales only when earned. Two weeks in the field revealed insights that no report could.
Designing for AI means designing for uncertainty. Every decision had to anticipate variability and make it feel human.
Design can define strategy. TravelGPT didn’t just improve UX — it redefined TBO’s roadmap for the future of travel.
Your screen’s great, but the desktop version is even better.
