Featured System98% Match20264K ULTRA

AI Concierge for a Regional Resort Group — 91% Self-Service, 321 Leads in 5 Weeks

An AI concierge live across Facebook Messenger, LINE, and the brand website for a hot-spring resort in a regional hotel group — handling 91.1% of guest conversations on its own and capturing 321 booking leads in its first 5 weeks.

Verified Case Results

Self-Service Rate

91.1%

Leads Captured

321

Uptime

99.4%

active-worker-sf-stream

LOG: PWA offline worker initialized cache index successfully.

All Case Studies

AI ChatbotAI Concierge for a Regional Resort Group — 91% Self-Service, 321 Leads in 5 Weeks

Live · 2026Case study

AI AutomationAI Agent for Automated Expense Entry — Import/Trading Business

Live · 2026Case study

AI ChatbotLINE AI Chatbot for an Industrial Valve Distributor — 90% Self-Service Across 3,260 SKUs

Live · 2026Case study

AI AutomationCross-Company Accounting Automation — Multi-Property Hospitality Group (Phase 1)

Phase 1 Live · 2026Case study

AI ChatbotLINE File-Memory Assistant — MVP Delivered 23 Days Early

Live · 2026Case study

AI Chatbot Case Study

AI Concierge for a Regional Resort Group — 91% Self-Service, 321 Leads in 5 Weeks

98% Match Rating•Live · 2026•AI Chatbot

Full Case Study

The Client

A hot-spring resort in Kanchanaburi, the first property in a larger regional hotel and resort group's multi-property AI rollout. Guest enquiries were arriving across Facebook Messenger, LINE, and the brand website, in multiple languages, with no single system tying them together.

The Challenge

•Guest messages spread across three separate channels, each answered manually

•Enquiries came in Thai, English, and other regional languages, with no consistent response quality across languages

•After-hours and peak-time messages went unanswered until staff were back online, and every unanswered message was a booking lead at risk of going to a competitor

The Solution

A conversational AI concierge was deployed across all three channels at once — Facebook Messenger, LINE OA, and the resort's own website widget — so guests get the same quality of answer no matter where they reach out. The bot handles common questions about rooms, rates, and availability, guides guests toward a booking, and hands off to a human staff member the moment a conversation needs judgment a bot shouldn't make. Built and shipped in 14 days as the pilot property for a larger multi-property rollout across the group.

Why a Pay-Per-Job Model, Not a Monthly Retainer

The most consequential decision on this project wasn't the bot itself — it was how the client would pay for support after launch. Leadership was offered two options: a fixed monthly retainer, or pay only for the system-level work actually used, with no minimum. They chose pay-per-job, locked before build started, because a hotel group running many properties sees different support workloads month to month, and a flat retainer would overpay in quiet months. The front-desk team can edit FAQs, pricing, and promotions themselves at no charge through a shared spreadsheet and doc — extra cost only applies to system-level work like bug fixes or tuning. A 90-day care window is included first (free bug fixes within 24-48 hours, weekly reporting, automated monitoring) before switching over to pay-per-job.

Key Technical Decisions

•One shared conversational core deployed identically across Messenger, LINE, and the website widget, rather than three separate bots, so guest experience doesn't depend on which channel they picked

•Multi-language support built in from day one across 4 languages, since the resort's guest base isn't Thai-only — for languages beyond the core set, the bot automatically falls back to English

•The bot's knowledge comes directly from the resort's real website, pulled with an automated scraping tool rather than hand-typed from scratch, so its answers match what the resort already communicates publicly

•Two years of real historical chat history (with identifying details stripped) were used to train the bot's tone to sound like the resort's actual front-desk team, not a generic script

•A clear handoff boundary: the bot escalates to a human the moment a question needs judgment rather than guessing, with a one-tap 30-minute pause link sent straight to the staff LINE group

•Only anonymized user reference IDs are stored, never real names, phone numbers, or emails, with a signed data protection agreement in place from day one

From Kickoff to Production in 14 Days

•Day 1: Kickoff — access opened across all 3 channels, the AI account structure set up, confidentiality and data-protection agreements signed

•Days 2-4: Knowledge gathering — 50 FAQs, room photos, and details on activities and attractions in the area

•Days 5-7: Build — historical chat tone analysis, the bot assembled across 4 languages, Rich Menu shortcuts installed, chat widget embedded on the website

•Days 8-10: 35 real-world test scenarios run across all 3 channels before the client saw it

•Days 11-13: UAT with the client's own team — tone adjustments and FAQ fixes based on real staff feedback

•Day 14: Full production launch across all 3 channels, with continuous monitoring through the first 24 hours

Before going live, the client's own team verified all 12 agreed acceptance criteria themselves — language accuracy, all 50 FAQs present, the Rich Menu working, leads writing correctly to the tracking sheet, handoffs behaving as designed, all 35 test scenarios passing, and confirmation they could operate the system themselves after training. Launch happened only once every criterion was independently confirmed — no shortcuts taken to hit the date.

The Results

•91.1% of guest conversations handled by the bot on its own in the most recent week measured, clearing both the month-1 (70%) and month-3 (85%) targets set at the start

•321 cumulative booking leads captured in the first 5 weeks live, climbing steadily from 80 to 237 to 321 with no dead weeks

•Roughly 99.4% uptime across the first 13 days, with zero conversations left unanswered (0% blank-reply rate)

•Now live in continuous production, with the rest of the group's properties queued to roll out next

Staying Transparent About What Didn't Go Smoothly

This project wasn't friction-free, and it's worth recording the rough patches honestly rather than hiding them:

•The bot went completely silent for 41 hours at one point, caused by a misconfigured message-receiving connection during a system edit — real guest messages were silently dropped even though the backend kept reporting "success" on every run. It took ruling out 9 wrong hypotheses before finding the real cause, then confirming the fix with live test messages across all 3 channels.

•The bot's core instruction set was accidentally overwritten during a live edit, causing it to run briefly without its trained tone or guardrails — caught quickly and restored from a backup copy, with a new process in place to prevent it happening again.

•Handoff-to-human rates on Facebook (25%) and LINE (21%) ran slightly above the 20% target, but every one of those handoffs turned out to be by design — a guest asking for a human, or a payment/cancellation case meant for staff — not the bot failing to answer.

•The self-service rate was mistakenly reported as a flat 100% for over a month due to a calculation bug (a timestamp column silently went blank whenever staff paused the bot, so every human intervention went uncounted). Caught and corrected, with historical numbers recalculated — the 91.1% figure in this case study is the corrected number.

Lessons Learned

•A system reporting "success" doesn't mean a message actually arrived. The 41-hour outage happened because a connection was routing messages to a dead end while still reporting success — nothing looked broken from the outside. Now, a fix only counts as done once verified with a real message that reaches its real destination, not just a status check.

•Keeping a bot's core instructions in only one place is a silent failure waiting to happen. After the instruction-wipe incident, every core prompt, knowledge base, and critical piece of code now keeps a dated backup copy in the project folder, separate from the tool actually running the system.

•A metric that looks too perfect for too long is a signal to investigate, not celebrate. A flat 100% self-service rate sat unquestioned for over a month before anyone checked the underlying data and found it was hiding a bug, not reflecting reality.

•"Fixed" has to mean verified working in production, not just that the code looks right. A support ticket was once marked resolved after only a code review, before any live message confirmed the actual reply — the standard changed so nothing is marked fixed until a real test message confirms it.

•Don't trust a plausible-sounding theory until it's proven against real data. Three reasonable explanations were tested and ruled out, one after another, before the actual root cause — the instruction-wipe above — was found.

Project Information

ROLEAI Systems Architect

CLIENT / PARTNERRegional Hotel Group (NDA)

RELEASE DATELive · 2026

INTERESTED?

TECHNOLOGY MANIFEST

LINEFacebook MessengerMulti-Languagen8nClaude

More Like This

AI Automation

AI Agent for Automated Expense Entry — Import/Trading Business

96% Match

An AI agent that reads, classifies, and enters roughly 10 expense documents a day into the client's PEAK accounting system — validated against 45 of 48 real documents (93.75%) before go-live, self-hosted with no ongoing platform fee.

Live · 2026Read Case →

AI Chatbot

LINE AI Chatbot for an Industrial Valve Distributor — 90% Self-Service Across 3,260 SKUs

96% Match

A LINE chatbot that answers pricing questions, drafts quotations and invoices, and reads product photos and payment slips for a one-person industrial valve and pump distributor — handling 90% of routine questions across a 3,260-item, 52-brand catalog, built in 7 days.

Live · 2026Read Case →

AI Automation

Cross-Company Accounting Automation — Multi-Property Hospitality Group (Phase 1)

96% Match

A cross-company accounting automation pipeline for a multi-property hospitality group's single accountant — pulling files from OneDrive, reconciling bookings against bank statements and accounting-software records, and sending a daily AI-written summary to LINE. Phase 1 passed all 9 planned test items in 10 days.

Phase 1 Live · 2026Read Case →

Secure TLS Port connected • Fully Keyboard navigable