Conversational Travel Agent
Planning a trip used to mean 5 websites, 3 hours, and a spreadsheet. We replaced all of it with a conversation.
Travel planning is legitimately broken. Users bounce between Kayak, Google Flights, TripAdvisor, Airbnb, Yelp, and a Reddit thread from 2019 — and still end up unsure if they made the right call. The problem isn't lack of options. It's too many of them.
Setting the Scene
The average traveler visited 5+ different websites to plan a single trip — flights on one, hotels on another, experiences on a third, logistics on a fourth. Decision fatigue set in early. Hours of research led to suboptimal choices because users couldn't hold all the variables in their heads at once. What they actually wanted was an expert — someone who knew their preferences, could synthesize options, and would just tell them what to book.
The Problem
No single platform covered the full travel planning journey. Every tool optimized for its own category — flights, hotels, experiences — but nobody owned the trip as a whole. Users were the integration layer, doing the synthesis work manually, burning time and arriving at mediocre decisions. The opportunity was to build the trusted advisor layer that the category was missing.
What made it hard
- →Integration complexity — connecting to flight APIs, hotel booking, and experience platforms was genuinely hard
- →Recommendation accuracy — one bad flight recommendation destroys user trust permanently in this category
- →Natural language variation — users express travel needs in wildly different ways
- →Preference modeling from conversation is imprecise — people often don't know what they want until they see options
- →Booking friction — actually completing purchases through multiple backend systems in a single flow
Travel planning isn't about finding options. It's about making the right decisions and discovering experiences. Users didn't want to compare 100 flights. They wanted an expert to say 'for your preferences, flight X is the one — here's why.' Positioning the agent as a trusted expert changed everything about the design.
What the Research Revealed
- →5+ websites for most trip plans — users were the integration layer between every platform and it was exhausting
- →Decision fatigue was the #1 pain point — not lack of options, but paralysis from having too many
- →Personalized recommendations were highly valued — users were willing to pay for insider knowledge they couldn't easily find themselves
- →Spontaneity was important: 40% of trip satisfaction came from unplanned discoveries the traveler wouldn't have found on their own
- →Cost optimization mattered but wasn't the primary goal — 'best value for the experience' beat 'cheapest option' every time in research
The Approach
I designed a conversational agent that learned preferences through dialogue, made curated personalized recommendations, and handled all booking logistics. The key architectural decision was hybrid: conversation for preference discovery and discussion, structured UI for selection and confirmation. Pure conversation for booking felt risky. Pure forms felt like every other travel site. The combination earned both trust and efficiency.
Preference discovery through conversation, not forms
Rather than an upfront preference questionnaire, the agent discovered what mattered through natural chat — 'How many days?' 'What's your style — adventure or relaxation?' 'Budget range?' Felt like talking to a travel agent, not filling out a form.
Curated 3-4 options with explicit reasoning
Not 100 flight results — three. 'Based on your budget, timeline, and preference for minimal layovers, these three best fit.' The curation was the product. Users trusted the agent's judgment more when they saw fewer, better options.
Serendipity as a first-class feature
The agent proactively suggested experiences, alternative destinations, and off-script ideas — playing the role of a well-traveled friend with local knowledge. These suggestions drove 40% of users' highest-satisfaction trip moments.
Inspiration conversation → clean booking interface
Conversation mode for discovery. Clean structured UI for confirmation and payment. The mode shift happened naturally once decisions were made — no jarring transitions.
Proving It Out
Beta tested with 500 real users making actual trip bookings over 2+ months. Measured planning time directly — recording how long users spent from 'I want to take a trip' to 'I've booked everything.' Tracked recommendation quality through post-trip satisfaction surveys. And monitored booking completion rate from first recommendation to final confirmation.
- →70% reduction in planning time — from an average of 4 hours to 30 minutes. That's not an incremental improvement, it's a category shift
- →User satisfaction with recommendations scored 8.2/10 — strong signal that the curation was actually working
- →40% increase in average trip value — users booked better experiences when an expert was guiding them
- →85% booking completion rate — high conversion from recommendation to actual confirmed booking
Beyond the Numbers
- →Users felt like they had a personal travel advisor for the first time — the positioning landed exactly as intended
- →Increased spontaneity and discovery — users found experiences they wouldn't have found on their own and were thrilled about it
- →Reduced planning stress significantly — this was actually the #1 outcome users reported in post-trip surveys
- →Higher trip satisfaction overall — discovering unexpected experiences drove satisfaction above what users had planned for
What I Took Away
Conversational agents work well for preference discovery but struggle with complex comparison and decision-making. The hybrid conversation + structured UI approach worked better than either alone. And serendipity is genuinely underrated as a design value — proactive 'you might also love this' suggestions were worth more than any optimization feature we built.
- →Travel preferences are personal and contextual — building accurate preference models from conversation was harder than expected and required significant iteration on training data
- →Personalization doesn't always require AI — good recommendations based on user segment were often as good as individual AI personalization, especially early on
- →Transparency in recommendations matters enormously — explaining why a flight was recommended increased trust more than the recommendation itself
- →I chose curation over comprehensiveness — 3-4 good options vs. 50+ to compare. That tradeoff traded choice anxiety for confident decisions, and it was absolutely correct
What's Coming Next
Trip matching to connect travelers with similar interests. Dynamic pricing insights showing when fares and rates typically drop. A travel community where users share trips and get recommendations from people with similar taste. Insurance integration with appropriate coverage suggestions. Post-trip engagement with photo sharing, review collection, and trip analytics.