How anpira's AI concierge works

The concierge is a virtual front desk that answers your guests inside their booking portal: the same link where they find their arrival details, house guide and extras. Guests install nothing, sign up for nothing and need no extra phone number: they open their link and ask.

It is available from weeks before arrival (30 days by default) until a few hours after checkout (12 by default), which is when questions actually happen. It replies in the language the guest writes in.

Live conversation with the guest runs on Anthropic models (the Claude family). Host-side tasks that require reading documents and images (for example, extracting your welcome guide PDF or a photo of your house-rules sign) use Google's Gemini, which is multimodal.

anpira does not train any model on your data or your guests' conversations: every reply is generated on the spot by querying the model with the context of your property and that specific booking.

The concierge works exclusively with your property's information: the fact sheet (WiFi, parking, how to get there), the house manual (how the AC, hob or rubbish works), the rules, the extras with their prices, and the frequent mix-ups you have taught it ("the blue door is the garage, not the entrance"). On top of that comes the booking context: dates, stage of the stay and check-in status.

• If a fact is not in its context, it does not make it up: it says so naturally and, when appropriate, escalates to the host.
• It never hands out access codes or door keys via chat. It does not have them: the code is shown to the guest in the Arrival section of their portal, shortly before check-in.
• It does not reveal data from other bookings, guests or properties, and it is built to ignore manipulation attempts (changing its role, extracting internal instructions).

So you never fill in a form from scratch, you can hand it what you already use with your guests: a welcome PDF, a photo of a rules sign, your listing text or the messages you already send on WhatsApp. anpira reads it and returns a proposed concierge fact sheet.

Nothing is saved blindly: you review the proposal and decide what gets applied. And by design, extraction never includes door codes or access keys, even if they appear in the document.

You can give your concierge direct instructions ("don't recommend restaurant X", "mention the small pool to families with kids") and it always follows them. There is one exception: anpira's global safety policy, which overrides any host rule. That policy includes never inventing facts, never revealing other bookings' information and never giving out access codes via chat.

The concierge has exactly three actions, all auditable and logged in your dashboard. None of them moves money and none modifies bookings:

• Log an extra request (early check-in, late checkout, extra cleaning, pet, cot) when the guest clearly asks for one. When applicable, the option to confirm and pay for it appears in the guest's portal.
• Log a real property incident (something broken, something not working) with its status: open, resolved or escalated.
• Alert the host when a human needs to step in.

And just as clearly, what it cannot do: it does not offer discounts, approve compensation, negotiate refunds, grant free early or late stays, modify bookings or promise repair schedules on your behalf.

Before alerting you, it tries to solve. Its internal rule is literal: faced with a problem, its job is not to hand it off, it is to help solve it. WiFi that will not connect (forget the network and re-enter the password), a guest who cannot get in ("are you at the blue door? That one is the garage"), a minor clog, an appliance that will not start: it walks the guest through it step by step, with common sense and the house manual, without inventing instructions for appliances it does not know. If it gets solved, it closes the matter kindly and it is logged as resolved: you find out, but you never had to step in.

The concierge's goal is for you to reply as little as possible, yet find out in time when it truly matters. It alerts you when money is involved (refunds, disputes, cancellations), when a fault needs physical intervention, in emergencies, and in one subtler case: it reads the guest's mood.

If it detects frustration it cannot defuse (the guest repeats the complaint, raises the tone, threatens a review), it does not wait for things to blow up: it alerts you early, without drama and without telling the guest, describing how the person is doing so you arrive with context. Meanwhile, it keeps attending.

When you take over an escalated conversation, the guest can tell: the concierge steps aside and your messages arrive as a person's, because they are.

• The host does not read the guest's raw conversation with the concierge. They see a per-booking summary: what the guest asked and how each item ended up. Incidents and requests do arrive in full, because they are operations.
• Each host's data lives isolated from everyone else's at the database level (tenant isolation with Row-Level Security).
• The chat has usage limits and abuse protection, and only works within the stay window.
• Personal data processing is governed by anpira's Privacy Policy (GDPR).

• It does not reply on WhatsApp. The chat lives in the guest portal: nothing to install, the booking is the credential, and the full context of the stay is available. Host alerts do reach the host where they already work.
• It does not touch money: no discounts, no compensation, no refunds. Extras are paid online through their own flow, confirmed by the guest.
• It does not state occupancy limits or allow use of the property outside the booking dates.
• It does not recommend other properties or platforms.
• It never blames anyone: not the cleaner, not maintenance, not the host. It focuses on solving.