How to Train an AI Concierge on Your Property: What to Load First

You train an AI concierge on your property by building a knowledge base it answers from. Start with the 30 questions your desk answers every day, then policies in their exact wording, then menus and activities as structured data, then local tips with a tone guide, and finally the list of things it must hand to a person. That order is the whole job. The software part takes an afternoon; the knowledge part takes about three weeks, and it is the part that decides whether the pilot works, which is why our guide to the best hotel AI concierge platforms treats the knowledge base as the product.

One clarification before anything else, because the word "training" causes most of the confusion. Nothing here means the model learns from your guests. It means your team writes down what the property knows, in a form the AI can answer from, and keeps it true.

What does "training an AI on your property" actually mean?

It means curation, not model training. The AI reads a guest's question, finds the entries in your knowledge base that apply, and writes an answer from them. The entries are documents your front office manager can open and fix. The model itself does not change, and in a well-built system guest conversations are not fed back into it.

That is the safer design. If the AI answers from entries, a wrong answer can be traced to an entry and fixed in one edit. If it answered from something it "learned," nobody could say why it said the pool closes at 8. The industry's caution here is well founded: the NYU SPS Tisch Center's State of Distribution 2026 report found trust, governance and data privacy concerns limiting how far hotels let AI act on its own ^[1]. A knowledge base you can read is the answer to those concerns. The broader distinction is covered in how an AI concierge differs from a chatbot; for this post, the point is simpler. The AI knows what you wrote down. Nothing more, and if it is built right, nothing less.

Why is the loading order the whole game?

Because a wrong answer about a rare policy hurts less than a wrong pool hour at 4pm. Load by guest impact and by how often the fact changes. The things guests ask 50 times a day go in first and get the most care; the things one guest a month asks go in later or go to a person.

Guests reward this. In Hotel Tech Report's 2025 survey of 402 hotel guests, 70% found chatbots helpful for simple tasks ^[2]. Simple and frequent is where the AI earns trust, and trust is what lets it handle the harder tiers later.

Loading order for a hotel AI knowledge base, 2026. Recommended practice; adapt owners to your org chart.

Tier one: what are the 30 questions your desk answers every day?

Do not guess. Pull them from the desk log and the last 200 guest messages, and count. The list will be boring, and that is the point: pool and breakfast hours, Wi-Fi, parking, checkout time, late checkout policy and price, how to get towels, the gym, the kids' club schedule, the shuttle, the nearest pharmacy, restaurant reservations, when the bar closes, how to reach housekeeping.

Write one plain answer per question with a date on anything that changes. "The main pool is open 7am to 9pm. Updated June 2026." Not a paragraph. Not a PDF. Not "see our website." ResortConcierge AI holds these as an editable property knowledge base the team owns and the concierge answers only from, so the pool-hours fix is one edit by one person, not a retraining project.

If you have ever traveled with a child who asks "what time does the pool open" four times before breakfast, you understand the value of an answer that is instant, correct and does not sigh. That is tier one.

Tier two: how do you load policies without creating commitments you did not make?

Exact wording, from the signed version, with a date and one owner, and a rule that the AI quotes policy but never decides whether an exception applies. The cancellation policy goes in word for word. "Can I cancel without a fee because my flight was delayed" does not get an answer from the base; it gets a person.

The reason is a case, not a theory. In February 2024 a British Columbia tribunal ordered Air Canada to pay a customer CAD 812 after its chatbot gave him incorrect bereavement fare information, rejecting the airline's argument that the bot was a separate legal entity: it was "still just a part of Air Canada's website" ^[3]. That is a Canadian tribunal decision about an airline, not legal precedent for hotels, and this is general information rather than legal advice. The lesson holds anyway. Three versions of a pet policy in a shared drive is not a knowledge base. Pick the current one, date it, and delete the others.

Tier three: how should menus go in?

As structured data: items with a name, description, price and prep time; modifiers as options rather than a notes field; hours per menu; and an allergen flag per item. Never as a PDF. A PDF menu on a QR code is a picture of a menu, and the AI cannot take an order or answer "what is gluten free after 9pm" from a picture.

In the United States the FASTER Act made sesame the ninth major food allergen effective January 1, 2023, joining milk, eggs, fish, shellfish, tree nuts, peanuts, wheat and soybeans ^[4], so those nine are the minimum flags per item, reviewed by the chef and not by marketing. Allergen rules vary by country; this is general information, not legal or food-safety advice. The AI discloses the flags. It never confirms that a dish is safe for a particular guest; that question goes to a person who asks the kitchen. ResortConcierge stores structured menus with hours and allergen flags, which is what lets the concierge answer from live data rather than a scan.

The full order path from the chat menu to the kitchen pass is in how room service orders move from chat to the kitchen.

Tier four: how do activities, spa and rentals work?

Each activity gets an availability window, a booking rule, limits (age, weather, group size) and a flag for whether the AI may book it or only describe it. "Paddleboarding, 9am to 4pm daily, weather permitting, ages 8 and up, 60 minute slots, book up to 24 hours ahead" is an entry the AI can act on. "Ask at the activities desk" is a sign, not an entry.

Decide the booking question deliberately. Low-stakes, well-defined items (a cabana, a kayak, a massage slot) can be booked in chat. Anything with a deposit, a medical form or a judgment call gets described by the AI and booked by a person. ResortConcierge treats activities with availability and booking rules as structured objects, so "is paddleboarding available Tuesday" is answered from the calendar rather than from a brochure that was true in March.

Tier five: how do you add local recommendations without sounding like a brochure?

Write a one-paragraph tone guide first, then five recommendations per category, each with distance, price range and one honest note. The tone guide says: plain, warm, no superlatives, say whether it is a walk or a cab, mention if a place is loud or slow or cash only. The honest note is what guests actually want: "excellent fish, 20 minute wait most nights, bring bug spray for the patio."

Categories for most resorts: restaurants by price and type, a pharmacy and a clinic, a grocery, kid-friendly outings, rainy-day options, other beaches, and a taxi note. Refresh quarterly. Nothing undermines trust faster than a confident recommendation for a place that closed in April.

Tier six: what should the AI never answer alone?

Medical questions, safety and security, complaints, money and billing, legal matters, and anything not in the base. This list is loaded last and matters most, because it becomes the escalation rules. Each item is a trigger: the AI acknowledges, says a person is coming, and moves the conversation to the queue.

The last item on the list is the quiet one. "Anything not in the base" means the AI's answer to an unknown question is "let me get someone," not a guess. ResortConcierge sends questions with no match in the knowledge base to the escalation queue through smart escalation rather than improvising, and that queue is also the best source of what to load next, because every escalation for "AI did not know" is a missing entry with a guest's name on it. The reasoning behind each category, and how to phrase refusals so guests thank you for them, is in what an AI concierge should refuse to answer.

How do you test before guests see it?

With 50 real questions from the desk log, scored three ways: right, wrong, or escalated correctly. Not 50 questions you made up, because you will make up the ones you already loaded. Pull them from last month's messages, typos included.

Testing scorecard, illustrative rows, 2026. Score 50 questions; fix every wrong answer at the entry, then rerun.

Illustrative timeline for a 150 room resort, one front office manager part-time with help from the chef and the activities lead: week one, tiers one and two; week two, tiers three to five; week three, tier six, the 50 question test and fixes. About three weeks. The software may be live in an afternoon. The knowledge is not.

Who owns the knowledge base, and how do you keep it true?

One owner overall and one per section, written down, with a five minute weekly review and a monthly read of the escalation log. Accuracy is not a launch property. It is a routine.

The ownership question is urgent because the knowledge currently lives in people, and people leave. In the AHLA and Hireology 2025 survey of 282 hoteliers, 65% reported staffing shortages and 71% had unfilled openings, averaging six to seven per property ^[5]. The two long-tenured staff who know why the north pool closes early on Thursdays are a single point of failure. A written base is how the property keeps what they know, and it doubles as a training document: using the same knowledge base to onboard seasonal staff means the AI and the new hire learn from the same page, and when one is wrong, both get fixed.

Scattered information is the other enemy. Access Hospitality's 2026 research found hotels using an average of five systems, with staff spending 78 minutes a day switching between them or combining data ^[6]. The hours live in one place, the menu in another, the policy in an email from 2024, and when property information lives in five systems, the knowledge base is where it finally gets reconciled. Seasonal changes are the test: put the summer menu date on the calendar, assign the owner, and make the edit the morning it changes.

Three capabilities to note as planned, not shipped, on ResortConcierge: PMS connectors for reservation data, a supervised mode for approving replies in the first weeks, and voice. Today, every entry is put there by your team, which is more work and also exactly why you can trust it.

Start tonight with the desk log. Thirty questions, one plain answer each, a date on every one. That list is tier one, it is most of what guests will ever ask, and once it exists the rest of the base has a shape to grow into.