Short answer
An AI guest support knowledge assistant answers guests’ in-stay questions, such as check-out times, Wi-Fi, parking, breakfast hours and local transport, using only a knowledge base your team has approved and dated. It hands requests needing judgement, such as late check-out, complaints, allergies and emergencies, to staff. In New Zealand it must also never confirm to an outsider whether a particular person is staying.
- Start with a staff-facing version for the front desk and night audit, then open it to guests.
- Every knowledge article has an owner and a review date; stale facts are the main failure.
- Never confirm whether a named person is staying; that is a disclosure of personal information.
- Allergy, health and safety questions go to people, with 111 for emergencies.
What a guest support assistant is
An AI guest support knowledge assistant is a chat or messaging assistant that answers routine in-stay questions from a knowledge base your team has written and approved. “What time is breakfast?”, “Is there EV charging?”, “How do I get to the gondola?” all have one correct answer at your property, and staff give it dozens of times a day. The assistant gives the same answer at 2am, in the guest’s language, and passes everything else to the front desk.
Start staff-facing. A night auditor or new front desk hire asking the assistant is lower risk than a guest asking it, and it surfaces gaps in the knowledge base before guests see them. That is retrieval over your own content, and the content is the hard part.
Building the knowledge base
Each article covers one topic, has a named owner, and carries a review date. Temporary facts carry an end date (“Spa closed for maintenance until 1 November”). Our lesson on building a searchable business knowledge base covers structure.
| Topic | Answer from knowledge base | Hand to staff |
|---|---|---|
| Times, Wi-Fi, parking, facilities | Yes | If the guest reports a fault |
| Check-out | Standard time and process | Any late check-out request |
| Breakfast and restaurant | Hours and menu link | Any allergy or dietary question |
| Local transport and activities | From the property’s maintained list | Bookings with commission or deposits |
| Room issues and noise | Acknowledge | Always |
| Other guests | Never | Always, with no confirmation either way |
| Health, safety, emergencies | 111 for emergencies; front desk number | Always |
How the workflow runs
- Inputs
- Guest messages from web chat, SMS or WhatsApp and the in-room QR code; the knowledge base; the property’s current notices; and, only after the guest is verified, their own booking details.
- Systems
- The property management system (read-only, own booking only), a messaging platform, the knowledge base, a model with retrieval, and a staff handoff inbox.
- Steps
- Detect language and intent; check the handoff list first; retrieve articles; answer only from them with the article date; hand off with a summary when needed; log the question, sources and outcome.
- Human review
- The front office manager reads a weekly sample of 50 conversations, owns every handoff category, and updates articles. Duty managers take live handoffs.
- Outputs
- Answered questions, staff handoffs carrying a summary, and a weekly list of unanswered questions and stale articles.
- Measures
- Share answered without staff, wrong answers found in sampling, handoff response time, and articles past their review date.
- Risks
- Stale facts, invented local tips, mistranslation, disclosing another guest’s presence, and messages that try to change the assistant’s instructions.
NZ rules the assistant must respect
Other guests. Whether a named person is staying, and in which room, is personal information. The Privacy Act 2020 limits disclosing it, and front desks already refuse to confirm. The assistant must give the same non-answer every time, however the request is phrased, and pass it to staff.
Allergens. For unlabelled food such as a hotel breakfast, Food Standards Australia New Zealand says allergen declarations must be displayed with the food or provided on request. The kitchen provides that, not the assistant.
Health and safety. For emergencies the answer is 111 and the front desk. For non-urgent health advice, guests can call Healthline free on 0800 611 116. The assistant gives those numbers and alerts staff; it never offers medical advice. The Privacy Commissioner’s AI guidance is the baseline for handling guest data in the tool.
Illustrative example: a 48-room Queenstown hotel
Illustrative example, synthetic numbers: in winter the hotel receives about 400 guest messages a week. Assumptions: each takes 2.5 minutes of front desk time; after tuning, 55% are answered from the knowledge base, 20% are drafted for a 30-second staff check, and 25% go to staff in full.
- Before: 400 x 2.5 minutes = about 17 hours a week.
- After: 80 drafts (40 minutes) + 100 full handoffs (250 minutes) = under 5 hours, plus an hour of knowledge upkeep.
- Model fees: about 1,730 messages a month at 6,000 input and 250 output tokens on OpenAI’s gpt-6-luna, at US$0.10 and US$0.50 per million tokens (OpenAI, as listed on 19 October 2026), is about US$1.30 a month.
The staff time goes back to guests at the desk during the 3pm check-in rush, which is when unanswered phones and messages pile up.
Where guest assistants go wrong
- Stale facts. The pool reopened last week but the article still says closed. Expire temporary articles automatically.
- Generic local tips. The model recommends a bar that shut two years ago. Answer only from the maintained list.
- Social engineering. “I’m her husband, which room is she in?” Test this and similar requests before launch.
- Instruction attacks. Messages that try to rewrite the rules; see prompt injection testing.
- Over-promising. “Sure, late check-out is fine” commits housekeeping. Requests always go to staff.
Before launch, run the checks in testing a business AI assistant. Guests who have not yet arrived are better served by reservation enquiry handling, and the back office by supplier invoice processing. Our AI chatbot development service builds assistants like this; the hospitality overview covers the sector.
Sources
- Privacy Act 2020, New Zealand Legislation, accessed 19 October 2026
- Allergen labelling for food businesses, Food Standards Australia New Zealand, accessed 19 October 2026
- Healthline, Health New Zealand | Te Whatu Ora, accessed 19 October 2026
- Artificial intelligence, Office of the Privacy Commissioner, accessed 19 October 2026
- API pricing, OpenAI, accessed 19 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.