Harborlight Hotel AI Concierge Agent
A Hong Kong boutique hotel deployed a multilingual AI concierge that answers 79% of guest queries before a human is paged.
Impact
What changed.
Coverage and language
79% of guest queries are now resolved by the agent without paging a human. Japanese-speaking guests report a step-change in service consistency.
Concierge focus
Human concierges now spend their time on high-value guest moments — bespoke itineraries, special occasions, complaint recovery — instead of breakfast-hours lookups.
Booking conversion
In-room dining and restaurant booking conversion rose 47% as the agent surfaces options in the guest's language, in their chat channel, in seconds.
The challenge
Before
Harborlight is a 38-key boutique property in Sheung Wan with a strong international guest mix. Front-desk and concierge staff were fielding roughly 240 guest queries a day in four languages — directions, restaurant recommendations, late check-out requests, laundry pickup, transport. The two-person concierge shift could not give every guest the unhurried attention the brand promised, and after-hours queries were either missed or escalated unnecessarily to the duty manager.
- 240 daily guest queries across four languages handled by two-person shifts
- Recommendations given quickly under pressure, often missing the better option
- After-hours queries either missed or escalated to the duty manager
- Restaurant reservations made by staff phoning around mid-conversation
- Repeated questions (wifi, breakfast, gym hours) consuming concierge bandwidth
- Multilingual coverage uneven — Japanese support depended on one staffer
- No record of what each guest had asked or been told
- Brand-voice consistency varied between staff members
The solution
What we built
We deployed a guest-facing concierge agent that lives in-room (via QR-code chat), in the hotel app, and on WhatsApp / WeChat / LINE for guests in their preferred channel. The agent answers in English, Cantonese, Mandarin, and Japanese with brand-trained voice. It handles factual queries directly (hours, amenities, transport), executes transactions through integrations (restaurant booking via Hong Kong booking platforms, in-room dining via the PMS F&B module, housekeeping requests routed to the duty queue), and escalates anything ambiguous or sensitive to a human concierge with full context attached. Every interaction is logged against the guest profile, so the next morning's concierge knows the guest asked about hiking trails the night before. Brand voice is consistent because the agent is trained on curated examples — recommendations match what the hotel actually wants to recommend, not generic TripAdvisor noise.
Core workflow connections
How the system flows.
- Guest QueryLanguage DetectionIntent Classification
- Factual AnswerBranded ResponseConversation Log
- Booking RequestIntegration CallConfirmationFolio Update
- In-room Service RequestDuty QueueStaff AcknowledgementClosure
- Ambiguous or SensitiveHuman Concierge Escalation with Full Context
- Curated RecommendationsBrand VoiceHotel-preferred Vendors
- Multichannel access via in-room QR, hotel app, WhatsApp, WeChat, LINE
- Guest profile memory carried across interactions and stays
- Conversation handoff to next-shift concierge with summarized context
- Quality review queue surfacing low-confidence agent responses for tuning
Process
How we built it.
Guest Query → Language Detection → Intent Classification
Factual Answer → Branded Response → Conversation Log
Booking Request → Integration Call → Confirmation → Folio Update
In-room Service Request → Duty Queue → Staff Acknowledgement → Closure
Ambiguous or Sensitive → Human Concierge Escalation with Full Context
Curated Recommendations → Brand Voice → Hotel-preferred Vendors
Multichannel access via in-room QR, hotel app, WhatsApp, WeChat, LINE
Guest profile memory carried across interactions and stays
Conversation handoff to next-shift concierge with summarized context
Quality review queue surfacing low-confidence agent responses for tuning
Start a project
Concierge team buried in repetitive queries?
We build concierge agents that protect your brand voice and your team's time — handling the easy 80% so humans can deliver the moments that matter.
No retainer lock-in · Month-to-month · Full transparency
More work
Up next.
A Singapore-headquartered regional insurer cut first-touch claims triage from 6 hours to under 8 minutes.
Sentinel Insurance Claims Triage AI
A Sydney RPO staffing AU and NZ clients processed 4,400 monthly CVs against 21 active reqs with an AI screener.