Concierj
AI-powered pre-booking guest communications & intelligence product for boutique properties — 0 → 1 solo build
Overview
Boutique hotels are losing potential guests before they ever book — not because of price or availability, but because of a gap in the pre-booking experience. No guided consideration stage exists, no system to answer questions in real time, qualify fit, or move an interested guest towards a booking decision. I built Concierj to solve that and become a property's personal travel agent - capturing interested guests and guiding them to booking.
AI-powered guest communication & intelligence product, solo build in 12 weeks, from concept to live launch.
I saw a gap and harnessed this opportunity to not only solve the gap, but to challenge myself to learn and deploy an AI product.
- 1Research
- 2Conceptual Design
- 3What I Built & Why
- 4Usability Testing
- 5Prototype
01 - Research
My research is rooted in human-centered design with the belief that end users should be included in all aspects of the process and product design. I conducted 59 interviews across four stakeholder groups - boutique hotel owners, front desk staff, guests, and travel agents to understand where the pre-booking breakdown actually happened and for whom. I wasn't looking to validate a solution; I was looking for the truth about the problem.
Alongside primary research, I analyzed the competitive landscape, assessed the total addressable market for AI-powered hospitality tools, and identified which data points would be most significant to track across the chatbot and the guest intelligence report.
02 - Conceptual Design
Research identified a missing consideration stage between website visit and booking. The next step was defining how guests would move through that experience, what information would be collected, and when human intervention was necessary.
As part of the architecture process, I evaluated multiple technical approaches and refined the solution to a streamlined Botpress and Claude stack, eliminating unnecessary complexity identified during the research phase. The resulting user flows, system architecture, business rules, and data requirements served as the blueprint for development of the functional prototype.
The Guest Journey Map
Following stakeholder and user interviews, I created a guest journey map to better understand how prospective guests navigate a hotel website during the decision-making process. The exercise helped further identify moments of friction, information gaps, and drop-off points within the traditional booking funnel, while revealing opportunities where Concierj could provide meaningful support.
One key insight emerged immediately: high purchase intent exists before booking, but guests often lack guidance during the consideration stage. As they compare accommodations and experiences and search for specific information, decision fatigue causes them to abandon the booking process altogether.
Key design decision: Built guided prompt bubbles that surface common guest questions and next steps. This approach reduces cognitive load, accelerates information discovery, and gently steers a guests towards information they are most likely seeking, creating a more intuitive and effective pre-booking experience.
Feature Map
I synthesized findings into a feature map to prioritize and define the MVP, identifying the ore capabilities needed to support guests during the hotel consideration stage while minimizing operational burden of staff.
Research didn't just inform the product, it changed it...4 times throughout the product build.
Low-Fidelity Systems Design
I began the systems design process with low-fidelity sketches to determine how different guest segments should be identified and routed through the experience, including lead qualification, escalation pathways, and human handoff triggers.
I designed end-to-end user journeys and conversation flows, mapping how prospective guests move from initial questions to booking decisions or human hand off via lead capture or escalation.
03 - What I Built and Why
Rather than building a simple FAQ chatbot, my goal was to create a system that could actively guide guests through the pre-booking journey while providing value to hotel staff behind the scenes. Each component was intentionally designed to support one of Concerj's three core pillars:
Conversational Workflow Architecture
What I built: Standard nodes (after hours, guest feedback, new v. returning guests) and autonomous node (decision tree for room & experience recommendations, lead capture, escalation paths).
Why: Interviews revealed guests wanted guidance, not FAQ answers. Needed a system that could handle common inquiries, escalate high-intent guests, and build on conversation.
Key design decision: The autonomous node handles 98% of the service, proactively guiding guests through trip planning using personalized recommendations and suggested conversation paths, while simultaneously capturing guest intelligence.
Knowledge Base Design
What I built: Structured hotel information repository including room data, amenities, dining, policies, and local experiences.
Why: AI quality is only as good as the information available and the guidelines in place to prevent hallucination. Early testing showed generic responses reduced trust.
Key design decision: Organize content around guest-decison making rather than internal hotel departments to improve AI-response time and reduce token usage.
Human Handoff & Email Notifications
What I built: Automated staff alerts, conversation summaries, and guest contact information.
Why: Hotels don't need more conversations — they need visibility into the conversations that matter.
Key design decision: Deliver actionable context rather than raw chat transcripts.
Guest Intelligence Report
What I built: Inquiry trend reporting, guest behavior insights, common questions, conversion indicators, and recommendations.
Why: Every guest conversation contains market research. Boutique hotels rarely capture or analyze that data.
Key design decision: Transform individual conversations into operational insights.
04 - Usability testing
I created a fully-functional, high-fidelity prototype of Concierj using Botpress. Throughout development, Concierj was iteratively tested through hundreds of real-world conversations and simulated guest scenarios across family members, friends, advisors, and prospective users. Testing ranged from structured booking scenarios to exploratory sessions designed to uncover edge cases, identify failure points, and evaluate the overall guest experience.
Finding #1: Guests weren't simply booking a hotel room — they were planning an expeirence.
During testing, I explored several approaches to helping guests begin their journey. While each improved information discovery, testing revealed that guests weren't simply looking for answers — they were looking for confidence in their booking decision. The final solution balanced both information retrieval and personalized guidance.
Iteration 1: Lead with Intent
The first iteration asked guests to self-identity their trip to mirror the kind of curated qualification a hotel concierge would do in-person. In testing, the approach surfaced a gap: travelers booking family trips, group travel, or simple logistics didn't see themselves in any option, and even guests with clear trip types, felt boxed into a single narrative before they'd had a chance to ask their actual question.
Iteration 2: Lead with Information
Next, I replaced the prompts with broader category bubbles - closer to a standard FAQ menu. Guests responded well to the directness: questions got answered fast, with no guesswork about where to click. But the format introduced its own friction. Without any framing around intent, guests comparing rooms or experiences across categories ran into decision fatigue.
Solution: Lead With Guidance
The winning decision combined the strengths of both: low-friction categories for guests who just needed information, alongside a new "Plan My Stay" option for guests who wanted guidance. Plan My Stay opens with a simple prompt - tell me about your trip - and the bot responds with a personalized, story-driven recommendation: specific rooms, activities, and special touches tailored to what the guest shared, helping them picture themselves at the property before any logistics enter the conversation. From there, guests naturally move into the practical questions already sold on the stay. Engagement jumped noticeably after the change, and Plan My Stay is now a core, defining feature of the Concierj experience.
Finding #2: Different Situations Require Different Summaries
Early versions of the summaries were accurate but inconsistent. Some included too much detail, creating noise, while others omitted important context needed for follow-up. Through iterative testing and refinement, i developed specialized summary frameworks for different conversation outcomes, ensuring hotel staff received concise, actionable information tailored to the next step required.
Key insight: The most useful summary depends on what action the hotel needs to take.
Finding #3: Real Guests Don't Follow Happy Paths
While the chatbot performed well during expected booking scenarios, testing revealed that users rarely follow ideal conversation paths. Participants frequently asked unexpected questions, changed topics mid-conversation, or interacted with the assistant in ways that differed from the designed workflow.
One of the most significant issues emerged when users opened the chatbot and immediately asked a question that was out of the assistant's scope. Rather than gracefully recovering, the assistant became trapped in an error state failed to continue the conversation, and failed to trigger the escalation workflow.
To improve resilience, I redesigned the conversation routing logic and introduced fallback handling throughout the workflow. When a question cannot be answered through predefined paths, the assistant now routes to an autonomous AI node capable of determining the best next action.
Additional refinements included:
- Expanding fallback handling
- Improved autonomous node instructions
- Enhanced escalation logic
- Dedicated "on failure" routing paths
The assistant become significantly more resilient, reducing dead ends and ensuring guests could continue receiving support even when the conversations moved outside predefined workflows. By introducing dedicated fallback and escalation pathways, the system was able to recover gracefully from uncertainty while minimalizing the likelihood of inaccurate or hallucinated responses.
Additional Iterations:
- Increased chatbot visibility through proactive welcome messages, notification indicators, and sound cues.
- Refined widget sizing and placement to improve discoverability across user groups.
- Adjusted onboarding copy based on user feedback.
05 - Prototype & Next Steps
The workflows and user journeys were initially designed in Figma before being implemented as a fully functional prototype in Botpress. Claude was used to generate and refine the custom code required throughout development, enabling rapid iteration and testing. The prototype underwent hundreds of conversational interactions and usability testing sessions, allowing me to continuously refine the guest experience, staff workflows, AI instructions, and conversation routing logic.
The result is a fully functional, AI-powered pre-booking guest communication system capable of guiding guests through the consideration stage, escalating high-value opportunities to staff, and transforming guest conversations into actionable intelligence.
Learnings
- The Real Problem is Rarely the Stated Problems: Research revealed a mismatch between what hotels were providing and what guests actually needed. It's important to say flexible in idea generation and to include the end users from the start.
- Speed Is a Design Tool: Building with Claude dramatically shortened the time between idea and prototype. The faster I could put a working experience in front of the users, the faster I could identify flawed assumptions, unconver edge cases, and iterate on the experience. Rapid prototyping allowed me to spend less time speculating and more time learning from real interactions.
- Designing AI Means Designing Human Workflows: This project reinforced that successful AI products must be designed around the people receiving the output as much as the people interacting with the system. Human workflows, handoffs, and decision-making processes are critical parts of the overall experience.
- Building a Product Requires More Than Design: What began as a simple side project evolved into a full-scale entrepreneurial venture. Beyond UX design, I found myself learning market positioning, product strategy, implementation decisions, technical constraints, and business viability. The experience reinforced that successful products are built through equal parts design, adaptability, and persistence.
The product is ready, so...what's next?
