Voyage: how much should a travel AI do on its own?
An AI travel concierge that plans, books and looks after a trip. The hard part was not making the AI capable. It was deciding, moment by moment, when it should act and when it should ask.
- Type
- Self-initiated AI product concept
- My role
- Product designer, UX AI engineer and product strategist
- Scope
- Plan, book and manage across 8 travel services
- When
- September 2026
- Stack
- Figma, Claude API, React, Express, Postgres, Vercel, Railway
- Status
- Live prototype on mock travel data
The autonomy ladder that shaped every AI feature. The more a mistake would cost the traveller, the higher the rung, and the more a person stays in control.
- Weather and advisory alerts
- Price drop notices
- Itinerary suggestions with reasons
- Destination match scores
- Low-risk rebooking after a delay
- Notifies with a one-tap alternative
- Multi-leg disruptions
- Visa, medical or family complications
- Payment and refunds
The problem
A single international trip touches flights, hotels, local transport, activities, insurance, currency, documents and, when things go wrong, rebooking. Travellers stitch this together across many apps and tabs, and the effort peaks at the worst moment: a cancelled flight at midnight in an unfamiliar airport.
AI travel assistants tend to fail in one of two ways. Some chat well but cannot act, so the traveller still does all the work. Others act too freely, and one wrong automatic booking destroys trust in everything else. I set out to design the space between: an assistant that does real work, with the limits of that work visible and predictable.
My role
I worked end to end as product designer, UX AI engineer and product strategist. I wrote the requirements and the scope options, designed the experience in Figma, built the prototype with Claude as both a coding partner and the in-product model, and defined the rules for what the in-product AI may do.
Every AI feature went through the same five questions before reaching a wireframe. Who benefits? What does a wrong answer cost? Can the traveller understand why? Can they override it? How does it fail?
Four decisions that shaped the product
Chat and search as equals, not chat-first
The first instinct with an AI product is to make everything a conversation. But travellers arrive in two states. Some have a vague wish, “somewhere warm for our honeymoon in December”, and benefit from conversation. Others know exactly what they want, a flight on a date, and a chat window only slows them down. Voyage gives both equal weight and keeps the trip context when the traveller switches between them.
Trade-off: two entry points mean more interface to design and maintain. A chat-only product would have been simpler and worse for anyone who already knew their dates.
The AI asks, the traveller chooses
Planning happens in a guided conversation. The concierge asks one thing at a time, with the right control inline: a calendar for dates, counters for travellers, option cards for flights, hotels, transfers and restaurants. Each card carries its price and the facts needed to choose. The traveller picks; the AI adds it to a running trip summary that shows every item and the total as the plan grows. When the AI finds something it could reuse, such as passports from a previous trip or loyalty points, it asks before using it.
Trade-off: the AI could assemble a whole trip from one message. Asking step by step is slower, but every item in the plan is one the traveller has seen and chosen.
Rebook automatically only when it cannot hurt
Disruption handling is where autonomy matters most. Waiting for the traveller to act wastes the minutes when good alternatives still exist. Acting on everything risks a rebooking that breaks a hotel night or a connection. Voyage is semi-automatic: it rebooks on its own only inside strict limits, asks when there is a trade-off, and hands complex cases to a person with the context already prepared.
Trade-off: strict limits mean the AI sometimes asks when it could have acted. In travel, one bad automatic rebooking costs more trust than many unnecessary questions.
| Situation | What Voyage does | Why |
|---|---|---|
| Delay or cancellation; same-day alternative on the same airline, no extra cost, no change to connections or hotel nights | Act, then tell | No downside to acting quickly; seats disappear fast |
| The alternative costs more, changes the day, or shifts a hotel check-in | Suggest | The traveller must weigh a trade-off only they can judge |
| Multi-leg trip, visa implications, travellers with children, older relatives or accessibility needs | Hand to a person | Too many dependencies for an automatic decision |
| Gate change, weather or advisory with no booking impact | Inform | Useful to know; nothing to decide |
Personalisation the traveller can see and edit
Voyage learns from past trips, and says so. The Travel DNA profile shows what it has learned, such as preferred season, typical trip length and budget, and who the traveller usually travels with, labelled “learned from 12 trips”. Beside it sit the controls: recalibrate, clear learning data, or switch AI learning off entirely in privacy settings. Stated preferences, such as travel style and amenities, are chosen by the traveller and the screen says exactly how the AI will use them.
Trade-off: showing what the AI has inferred exposes its mistakes. That is the point. A wrong inference the traveller can see is one they can correct or clear.
Where I overrode the AI
I built Voyage with AI tools, and they had opinions about the product too. These are the moments I said no.
Chat-only interface
AI proposed: a single conversational entry point.
I kept search equal to chat, because precise travellers should not have to explain themselves to a bot.
One-step itinerary
AI proposed: generate and fill the whole itinerary at once.
I made the concierge offer options step by step, so the traveller chooses every item in the trip.
Suggestions without reasons
AI proposed: cards with title, price and rating only.
I made a reason a required field in the response, so no suggestion appears unexplained.
Auto-confirm booking
AI proposed: a one-click “book everything” step.
One action books the trip, but only at the end, with every item and the total visible throughout. No money moves without an explicit confirmation.
Outcome
Voyage is a live prototype on mock travel data. It is a concept, not a launched product, so I separate what exists from what is still open.
What exists
- A working prototype covering home, destination discovery with AI match scores, the concierge conversation, a seven-step booking flow, the Real-Time Assistant and the Travel DNA profile
- Visible AI disclosures in the product, such as “AI tip” reasons on recommendations and “AI will” statements before automatic actions
- A written autonomy model and disruption rules
- Requirements and three scoped delivery paths with costs and trade-offs
Still open
- Testing with real travellers, starting with the disruption rules
- Live supplier integrations in place of mock data
- Calibrating the auto-rebook limits against real airline data
- Measuring whether visible personalisation builds or reduces trust
If you are building an AI product
The autonomy ladder is not specific to travel. Any AI product that can take actions on someone’s behalf needs the same four decisions made explicitly, feature by feature:
- What may the AI only tell the user?
- What may it propose but never apply?
- What may it do on its own, within which limits, and how does the user undo it?
- When must it stop and hand over to a person, and what context does it pass on?
Most AI products get the model right and the experience wrong. This is the layer I design.