Designing an AI travel agent that plans trips like a human advisor
fora is a conversational AI assistant that guides users from vague ideas to bookable plans — balancing intelligent automation with human control at every step.
Trip planning is cognitively overwhelming
Planning a trip means juggling budgets, dates, preferences, group dynamics, and logistics — often across 5 to 10 different tabs and apps simultaneously. The promise of AI is to consolidate all of that into one place. But most AI travel tools either feel too rigid (like a form pretending to be a conversation) or too loose (like a chatbot that suggests destinations without helping you decide).
The opportunity wasn't to build another search engine with a chat layer on top. It was to design something that feels like talking to a knowledgeable travel advisor — someone who listens, asks the right questions at the right time, and earns your trust before asking for your money.
How might we design an AI travel assistant that guides users from vague ideas to a concrete, bookable plan — without feeling like it's doing all the thinking for them?
What users actually do — and where they break
I spoke with eight people who had planned a trip in the last six months. The goal wasn't to validate assumptions — it was to challenge them. I also ran a competitive analysis across Google Flights, ChatGPT, Booking.com, and a few niche AI travel tools to map where each succeeded and where the gaps were.
The biggest assumption I came in with was that users want the AI to just tell them where to go. Research demolished that idea almost immediately. Users consistently wanted to feel in control — they wanted the AI as a sounding board, not a decision-maker. The moment they felt the system was choosing for them without explanation, trust collapsed.
From vague intent to a bookable plan
The conversation flow is the backbone of the entire experience. Every question the AI asks, every moment it pauses, every way it surfaces information — all of it had to feel deliberate and human, while still functioning as a structured planning tool underneath.
I mapped the core happy path and identified five critical design moments: the opening (validate before interrogating), the questions (choices over open fields, vibes over categories), the loading state (an honest pause builds credibility), the recommendations (framing before the cards), and the pushback (pivot gracefully, don't reset).
I also added a voice interaction mode — an alternative input channel where the AI listens, processes, and responds through a mic interface with a live waveform visualizer instead of typed text.
The voice mode presented its own set of challenges. A mic-based interface strips away all the visual scaffolding that text relies on — chips, cards, and inline suggestions. The design question became: how do you keep a user oriented when the only feedback is a waveform and a label?
The answer was clarity of state: idle, listening, and speaking each has distinct visual and color signals, with the waveform itself acting as the primary feedback channel. Green for listening, amber for speaking, grey for idle.
Earning confidence at every step
Trust is the most fragile thing in an AI-driven experience. It can be built slowly through small, consistent signals — and destroyed in a single moment of opacity. The design challenge here wasn't about adding a "why?" button. It was about making the AI's reasoning always visible, always honest — even when the recommendation isn't perfect. Toggle confidence levels · Explore reasoning + caveats
Keeping the human in the loop
An AI that can't be corrected isn't an assistant — it's a one-way street. The correction system is where the experience either feels like a conversation or like a form reset. The design principle I kept returning to: correct only what was challenged. Preserve everything else.
Four correction types were designed and tested. Price pushback — the most common — only updates budget and destination while vibe and dates hold. Vibe shifts are deeper: they reset the destination entirely but preserve everything else. Date changes are lightweight — the AI confirms and moves on without re-asking. And additive information (like "my friend might join") is handled not as a problem to solve, but as a new path to offer.
The live state panel on the right side of the prototype makes the correction logic visible — it's the core of what makes this section compelling. You can see exactly which fields the AI updated and which it left alone. That transparency is the same principle from the trust section, applied to the correction layer.
What I learned & what I'd do next
What I'd explore next: Multi-turn memory — how does the AI remember preferences across separate planning sessions without feeling surveillance-like? Error recovery — what happens when the AI gets something meaningfully wrong, not just imprecise? And the booking handoff in more detail — the moment the experience shifts from conversational to transactional is still the weakest seam in the flow.