Demo Observations
Six findings from the live demo, severity-coded by impact.
INDICATOR
Thinking Status
Current implementation is naive: renders a generic typing status, not reflective of actual agent progress. Not context-aware—same indicator whether searching flights or checking memory.
LATENCY
Flight Search
~42 seconds to find a route. Checkout flow has known bugs (long listing, layout issues). Economy auto-selected without explanation—should surface reasoning.
FUNNEL
Hotel Search
Three levels deep: results → profile → rooms → checkout. Profile page is LLM-generated, acknowledged as needing rethinking. Map view loads slowly. Gallery and room views functional.
PERSONALIZATION
Memory System
~10 minute refresh lag between conversations. Retains preferences but doesn't explain why ("you booked it 3 times" is the reason, but UI doesn't say that). Incorrectly retained emojis. Cold-start problem: no value until ~5th session.
CONTINUITY
Conversation & Trips
Each trip has one conversation. Draft trips exist (date + destination, no booking) but aren't surfaced. Farid: "My flight is still a trip whether I paid or not." Navan agreed but notes infra constraints.
NAVIGATION
Lost in the Homepage
Major known issue: users navigate away from a conversation and end up on the homepage with no clear way back. Current band-aid: "continue where you left off" button. ChatGPT/Perplexity sidestep this by having no homepage.
Trip Lifecycle
Navan breaks trips into four phases. Each has different information needs, different UI priorities, and different communication strategies. Active trips surface on the home screen with contextually relevant information based on timing.
Homepage & Component Surfacing
Farid raised the key architectural question: Can components live outside the chat? Navan's answer opens the biggest design opportunity in the engagement.
CONFIRMED
"Control of Everything Is Here"
Navan confirmed full control over rendering and updating components anywhere—not just in chat. Homepage, notifications, trip surfaces, wherever. Components can be maintained and updated for a user even when they're not in a conversation.
EXAMPLE
Price Hold Countdown
A "where you left off" component on the home screen showing an unbought flight with a countdown timer before the price hold expires. Drives re-engagement and urgency without requiring the user to re-enter the chat.
EXAMPLE
On-Trip Quick Links
During a trip, surface hotel address with direct link to Google Maps or Uber on the homepage. No typing needed. Active trips appear at the top with the most relevant information for that moment.
"Control of everything is here."
— Navan engineering, on rendering components outside the chat
Proactive Communication
Navan's vision is a "white glove" proactive service—an EA that handles disruptions before you even notice them.
Your flight was changed. We already rebooked you on the next one. Want to change it?— Navan's target experience for disruption handling
Loyalty & Rewards
The loyalty wallet aggregates all travel loyalty cards in one place. The challenge: it's a standalone feature, not integrated into the booking flow.
Loyalty Wallet
All loyalty cards in one place (United status, numbers, etc.). Users can jump from wallet directly into a conversation with loyalty context pre-loaded. Prompt examples on cards drive engagement—"a lot of people are trying those."
Standalone, needs integration
Navan Rewards
5% back on bookings, redeemable as Amazon gift cards. Rebranding from "Navan Rewards" to just "Amazon gift card"—that's what users actually want. Rewards are per-user total.
Rebranding in progress
Onboarding & Cold Start
The biggest open question in the engagement. Current approach: lazy onboarding—users don't see personalization value until ~5th session. Most never get there. Navan wants us to push the envelope.
CURRENT
Lazy Onboarding
Learn preferences over time through usage. Memory builds up across sessions. Problem: value compounds late. Early sessions feel generic. Users compare to Booking.com and don't see the difference.
REFERENCE
Lightweight Explicit
Ask a few questions upfront to personalize immediately. Farid cited Dia (by The Browser Company, acquired by Atlassian) as a strong reference. Simple mini-onboarding that personalizes the first steps.
OPTION
Gamified Onboarding
Pre-populated, quick, playful preference collection. More engaging than a form, faster than conversational. Could hit the ground running with travel preferences, loyalty programs, and style.
OPTION
Conversational Onboarding
Agent asks questions as part of the first conversation. Similar to Claude Code's approach. Natural but slower. Risk: feels like more "homework" if not done carefully.
Conversation Design
Six open questions Navan wants the UX team to take a stab at. These shape the entire conversational experience.
Tone of Voice
What register? How formal? The agent currently lacks a defined personality. Should it feel like a concierge, a friend, a professional assistant, or something else entirely?
Response Length
Short iterative exchanges vs. rich detailed responses? Current hotel results in 1.0 are an overwhelming "huge dump." How much is too much? How little is too little?
Pacing & Iteration
Fast iteration loops vs. comprehensive single responses? Do we want quick back-and-forth or thorough answers? The ~42s flight search latency constrains how iterative flights can feel.
Discovery Mode
How to help users who don't know exactly what they want? The happy path (known destination, known dates) works. Discovery mode—where the value proposition is strongest—is harder to design in conversation.
Information Density
Current hotel results in 1.0 overwhelm: a massive dump of UI elements and conversation. Farid: "You give me homework." How to restructure information delivery to feel like help, not work?
Education & Differentiation
Users comparing Edge to Booking.com don't understand the differentiation. The team acknowledged the gap in communicating value. How do we show value, not just deliver results?
You give me homework as a user. I was supposed to be helped, and now I need to work.
— Farid, on the current hotel search experience
Complexity & Edge Cases
Navan emphasized: the UX team must account for non-happy-path flows. The simple demo hides serious complexity underneath.
UNSOLVED
Round-Trip Modifications
"The minute you want to start making changes or reconsider or replan, it gets messy." Changing just the return flight while keeping the outbound is a conversation design nightmare. Not solved in 2.0.
COMPLEX
Multi-City Flights
Multiple legs, different dates, different constraints. The conversational interface must manage state across many decisions without overwhelming the user. Acknowledged as a major design challenge.
DEPTH
Hotel Funnel Levels
2 mandatory levels (hotel → room), optionally 3. Each level is a navigation step that can lose the user. The conversation must thread context across these levels without forcing restarts.
PRECEDENT
Farid's Lastminute.com
Farid noted prior experience designing for Lastminute.com (~2001) with exactly these problems. His position: don't reinvent the wheel—build on Kayak, Google Flights best practices while making the conversational layer feel magical.