Building an AI Copilot for Travel Consultants

From phone calls to first-draft itineraries

Track

Track 2 — Building and Implementing AI Tools & Agents in Production

Session type

30 mins talk

Abstract

At Travelopia, our team built TC Copilot for Enchanting Travels, a tailor-made travel company where Travel Consultants create highly personalized itineraries for guests.

Traditionally, a guest submits a lead, a Travel Consultant speaks with them to understand their preferences, and then manually creates a detailed proposal with cities, hotels, activities, transfers, flights, and other travel components.

When we started building TC Copilot, the problem looked simple: consultants were spending too much time creating itineraries, and we had years of confirmed proposals plus guest call transcripts. So our first version used a traditional RAG system over historical proposals, passed in a guest call transcript, and generated an itinerary.

Technically, it worked. The system produced a proposal.

But when we showed it to Travel Consultants, they rejected the quality.

That became the real learning. In tailor-made travel, an itinerary is not just a list of places and hotels. It needs to match guest intent, travel pace, accessibility needs, rooming preferences, hotel style, activity choices, and the kind of experience the business would confidently sell.

This talk shares how we moved from a demo-like RAG prototype to a more production-oriented AI copilot. We rebuilt the knowledge layer using GraphRAG, modelling cities, hotels, activities, routes, travel legs, styles, and preferences as connected data.

We also stopped asking the LLM to generate the entire trip in one shot. Instead, we broke the workflow into bounded steps: extracting requirements, building the route, selecting hotels, choosing activities, planning travel, and assembling a first draft.

The biggest shift was not just technical. We stopped framing AI as a replacement for Travel Consultants and reframed it as a copilot that creates an acceptable first draft. The consultant remains in the loop to review, curate, and apply final judgment.

The talk is about what it takes to move from “AI generated something” to “AI helped experts do real work faster.”

Why this talk is relevant

Many AI product journeys start with historical data, RAG, and a large prompt. The demo works, but production exposes issues around quality, latency, retrieval accuracy, domain judgment, and user trust.

This case study is useful for teams building AI tools in expert-led, high-context workflows where correctness is partly subjective and domain experts are the first real evaluation system.

Intended audience

Engineers, product managers, data teams, and technology leaders building LLM, RAG, GraphRAG, or agentic systems inside real business workflows.

Key takeaways

• Generating output is not the same as generating usable output.
• Domain experts are your first evaluation system before formal evals exist.
• Traditional RAG struggles when relationships matter more than isolated chunks.
• GraphRAG helped us query business knowledge more precisely.
• Breaking one large generation task into bounded steps improved speed and control.
• Human-in-the-loop is not a fallback; in expert-led workflows, it can be the product architecture.
• The goal changed from generating the final itinerary to creating a strong first draft that saves consultants time.

Talk outline

1. The traditional Travel Consultant workflow: lead, call, requirement capture, manual itinerary creation.
2. The first RAG-based prototype using sold proposals and call transcripts.
3. Why the demo looked successful to engineering but failed with consultants.
4. Reframing TC Copilot from replacement to first-draft assistant.
5. Moving from traditional RAG to GraphRAG.
6. Breaking itinerary creation into bounded steps.
7. What improved, what still needed human judgment, and why that was acceptable.
8. Lessons for building AI copilots in production.

Draft slides

WIP