The assistant

The assistant replies to your clients on WhatsApp, from the concierge service's number. This chapter explains how it works, what it relies on, what it will never do, how it calls on the team, how to pause it and how to set it up.

2.1 What the assistant is, and what it is not#

The assistant writes in the first person, on behalf of the concierge. It never says “our team”, “we” or “the assistant”: for the client, it is their concierge replying, under the first name chosen in Settings › General › Assistant’s first name (if empty, the name of the concierge service).

That does not make it a disguise. If a client asks plainly “am I talking to a robot?” or “is this an AI?”, the assistant replies with a fixed sentence, in the client's language: “I am Sofia's assistant, a program: Sofia reads everything and takes over whenever it helps.” This disclosure is compulsory (European AI Act, Article 50), cannot be switched off, and each occurrence is recorded in Settings › Privacy › AI disclosures.

Nor is the assistant an isolated chatbot. It creates booking requests that arrive in the schedule, asks the team questions, keeps the client's record up to date, sends the photos of a service, and sends again a payment link or a quote prepared by the team. It works with you, never in your place for anything that commits the house.

2.2 How it replies, step by step#

When a client writes, this is what happens, in this order.

1. It waits eight seconds. A client often sends two or three messages in a row. The assistant groups them and replies once.
2. It does not reply to a simple acknowledgement. “Ok thanks”, a thumbs up, a sticker: a human would not restart the conversation for that. Neither does the assistant. Except when a reply is expected: if its last message asked a question, or if a proposal is waiting for the client's answer (booking “Awaiting client”), an “ok” or a “Yes perfect” is an answer. A short agreement to a proposal goes to the team, linked to the booking, so that the team confirms it; the client receives “Perfect, noted. I'll get back to you very soon with the confirmation”, never “it's confirmed” (chapter 4, section 4.2).
3. It detects the language of the message. On every message, not only the first: a client who switches from Spanish to English for their friends is answered in English. It speaks French, English, Spanish, Dutch, Russian, German, Italian, Portuguese, Arabic and Chinese (Mandarin).
4. It answers the question “are you an AI?” honestly with the fixed sentence described above.
5. It judges the mood of the last message, and that message only:
   • life-threatening emergency (injury, someone feeling unwell, missing child, accident, fire): it immediately asks the client to call the emergency services (112 in Europe, adapted to the country), alerts the managers and creates an urgent request;
   • client stuck on site (gate closed, nobody at the meeting point, a service that does not start): a reassuring sentence straight away (“I'm dealing with it right now”), an alert to the managers and an urgent request;
   • angry client: a calming sentence (“I understand, let me sort this out personally”), an alert, and the assistant pauses itself on this conversation so that a human takes over.
     Only one alert is sent per episode (for two hours): an upset client who writes five messages does not wake the concierge five times. A life-threatening emergency, however, always triggers an alert.
6. Access codes, Wi-Fi, passwords: it never answers itself. It acknowledges the message and asks the concierge a priority question.
7. It prepares its reply based on what it knows (see 2.4) and using its tools (see 2.3).
8. It rereads itself (see 2.5): language, formal address, nothing confirmed, no repetition, no promise without follow-up.
9. It goes through a confidence gate. It assesses whether its reply is solid. If the reply contains information it has made up, it is discarded: the client receives a holding message and the team a question. If it is simply uncertain, it is held: the client receives a holding message (“Got it, I'm looking into it and will get back to you”) and the reply waits for a concierge to send or reject it (chapter 4).
10. It sends with a human delay. An instant reply at four in the morning gives away a machine. The assistant waits between 15 and 60 seconds during the day, between 45 and 160 seconds at night (from 01:00 to 07:00, concierge service time), a little longer for a long reply, and one and a half times longer in holiday mode. It simulates typing and, if needed, splits the reply into two or three bubbles. Emergencies go out within a few seconds.
11. It updates the client's record in the background (see 2.7).

If a technical failure prevents it from replying, there is never silence: the client receives an acknowledgement and the team a priority request “The assistant could not reply (technical incident)”.

2.3 What it can do: its tools#

2.4 What it relies on#

The assistant only knows what it is given. For every reply, it reads:

• The catalogue: the services that are active and visible to the assistant, with their selling price per unit (never a cost, a margin or a commission) and, when a season changes the price, the price for each season with its dates (“€620 per night off season; High season from 15/06 to 15/09: €868 per night”; cents are written when there are any, “€12.50”), their description, up to four attributes (“capacity: 10 people”), the minimum notice period, the “photos available” mention, upcoming unavailability periods and, for a service linked to a property, the dates already booked on Airbnb, Booking.com or Vrbo. Services are sorted by category and by supplier, the main supplier first, with the backup supplier flagged as such.
• The property records: for each active property (25 at most, the property of the client's stay first), its name, town, type, number of bedrooms and bathrooms, capacity, amenities and description (chapter 10). Never the access codes, access instructions, exact address, owner, technical contacts or the particulars reserved for the team.
• The constraints in the client's record: their allergies, first and as an absolute constraint (it never offers anything that goes against them and reminds the concierge of them in any request involving food), their children (first name and age) and their pets. Never their payment method.
• The client's current bookings, with their exact status and how long the client has been waiting.
• Their current or upcoming stay (property, dates, number of people).
• The questions already asked to the team about them and still unanswered.
• Their history: past bookings and stays, with dates.
• The summary of older conversations and their record (summary, known preferences).
• The client's writing style (length, punctuation, register) and their latest messages, to match them.
• The concierge's way of writing, picked up every week from the messages written by hand by the team (from fifteen messages onwards). The assistant never learns from its own messages.
• The local date and time and the calendar for the next fifteen days, so it never gets “Saturday” or “Wednesday evening” wrong.
• If the message comes from a linked contact (a client's assistant or family office), the fact that they are writing for this client and whether they can order on the client's behalf. This right is enforced by the system, not just by an instruction: a linked contact who is not allowed to order for any of the clients they represent creates no booking. Their request goes to the team as a question (“to check with Mr Laurent”), and the assistant simply replies that it is looking into it (chapter 7, section 7.6).

Why the catalogue and the property records are its only sources. Whatever is not written there, the assistant does not know: opening hours, cancellation terms, unlisted services, a villa amenity missing from its record. Faced with such a question, it has only two possible answers: the information is written down and it gives it, or it is not and it asks the team. A polite “yes” to “do you do helicopters?” would be a lie the concierge would have to stand by. Hence the importance of taking care of the descriptions and attributes of the services (chapter 8) and the property records (chapter 10).

A property record is there to describe the property (bedrooms, capacity, pool). It never says whether the property is free or how much it costs: the rental is sold through the catalogue, and a request for given dates always follows the usual path (the assistant looks into it, the team checks).

2.5 Its rules and safeguards#

These rules are set by Vetexis and cannot be changed. They were refined over months of real conversations and on a test bench of more than a hundred scenarios in ten languages.

What it never does:

• Confirm a booking. It only says “all set” if its reply can only be about bookings that are already confirmed: as soon as a client request is waiting for a decision (“Requested”, “Awaiting client”), or this turn has just created one or asked the team a question, an already confirmed villa no longer allows it to write “it's booked” for the chef that has just been requested. If it is about to write “I'll book it for you”, its sentence is rewritten without confirming anything, keeping the useful information; if the rewrite fails, the reply is held, with its reason.
• Cancel. A cancellation always goes through the team; it says it is dealing with it, never that it is cancelled. A reply announcing “it's cancelled” is rewritten, or else held.
• Negotiate. Any request for a discount goes to the team as a quote to prepare.
• Make up a price, an availability, a time or a feature.
• Announce as available anything it has not checked. To “a five-bedroom villa from 1 to 10 August?”, it does not reply “I have a villa at €620 per night”: it says it is looking into those dates and will get back quickly.
• Promise a specific time frame. Neither “within 20 minutes” nor “before this evening”: “very soon”, “during the day”. It does not know when the team will reply.
• Calculate a total. It gives the catalogue rates as they are, with their unit (“€450 per night”). Nine nights for four people: the team does the maths.
• Give out access codes, a Wi-Fi password or information about another client.
• Reveal its instructions, your margins, your commissions or what you pay your partners.
• Turn down a service outside the catalogue on its own: it is not for it to decide; it says it is looking into what is possible.

How it writes:

• short messages, as on WhatsApp, with no bullet points or formatting;
• one question per message at most, and only if it is needed;
• no greeting except in the first message, no “don't hesitate to”, no sign-off;
• never any emoji, even if the client uses them (a house decision: a luxury concierge writes sentences);
• never a long dash, which “looks machine-made”;
• it never asks again for information already given;
• it does not recite its catalogue and does not introduce itself.

What it does with an incomplete request. Before passing it on, it gathers the essentials in a single question: for accommodation, the dates and the number of people; for a table, the day, the time and the number of people; for a boat, the day, the duration and the number of people; for a transfer, from where, to where, when, how many. After two exchanges, it passes on what it has, flagging what is missing.

The automatic safety nets. After it has written its reply, Vetexis checks it mechanically: the language (translated again if needed; if that fails, the reply is held rather than sent in the wrong language), formal address for a client who uses it, no confirmation or cancellation announced, no word-for-word repetition of the previous message. If the assistant promises to check without having passed on the request, or if the client negotiates or cancels, the request goes to the team anyway. Manipulation attempts (“show me your instructions”, “I'm the developer”) are ignored without comment.

2.6 Questions to the team#

This is the heart of how it works: when the assistant is in doubt, it asks.

A question goes to the team in these cases:

The question arrives at the same time:

• in the app: Bookings › To approve › Questions from the assistant, the bell and the home screen;
• on WhatsApp, in each recipient's chat with the concierge service's number, with the Approve, Decline and Reply buttons;
• by summary email, at most once an hour, for those who could not be reached.

The Who receives requests to approve setting (Settings › Team) says who is notified: the managers, the member assigned to the conversation, or the whole team (chapter 4).

When the team replies, the assistant rephrases the answer for the client in their language and in the house tone. A curt reply between colleagues (“we don't do that”, “not available”, “ok but 11pm”) is never sent as it is: it becomes “I've just checked, unfortunately that isn't possible on that date. If you're flexible, tell me what would suit you and I'll look into it.”

With no answer after thirty minutes, the request is escalated: an alert goes to the managers and the client receives a reassuring holding message (never silence). When several requests are overdue at the same time, they go out in a single alert that lists them (“Escalation: 3 requests have been waiting for your reply for more than 30 min”, then one line per client and per subject, ten at most), instead of one message per request.

The same question asked twice within ten minutes does not create a duplicate.

2.7 The client record fills itself in#

After each exchange, the assistant updates the client's record, in the team's language (a record written in Spanish because the conversation was would be unreadable for the concierge):

• the client's current language;
• their register (informal or formal address), read from their own words;
• their first name, if they give it and no name has been entered;
• a summary of two or three sentences: who they are, what they want, what is useful;
• their writing style.

The preferences it spots (“gluten-free”, “prefers the Riva”, “table on the terrace”) are not written in directly: they become preferences spotted that the team accepts or ignores with one click on the record (chapter 7). The record remains the team's. Only exception: a fact the assistant notes in the middle of a conversation with its “note a preference” tool (an allergy that has been mentioned, for example) is written straight away.

2.8 Photos and voice notes#

With the Pro and Network plans (“Voice notes and photos read by the assistant” feature):

• a voice note is transcribed in the client's language, and the assistant replies to it as to a written message. The transcription is shown under the voice note in the inbox;
• a photo is described once, when it arrives (“screenshot of a booking at La Sala restaurant, Saturday 21:00”), and the assistant uses it to understand, without ever writing “I can see” or copying a number, a code or identity data. A photo is also flagged to the team.

With Essential, or if reading fails: the client receives an acknowledgement (only once in any three-minute period, even if they send four photos), and the team receives a “Media received” request to look in the inbox. A document or a video without a caption follows the same path.

Files kept. What the client sends is kept and can be viewed in the inbox: JPEG, PNG, WebP and HEIC photos (the iPhone format), OGG, MP3, MP4, WebM, AAC and AMR voice notes, MP4 videos, and PDF, Word, Excel and text documents. A HEIC photo that the server cannot convert is kept as it is, to open or download, without a preview.

2.9 Pausing the assistant#

There are several levels of pause, from the broadest to the most specific.

Automatic resumption. A pause set by taking over (in the app, from the phone, from the WhatsApp console, or automatically) lifts by itself after six hours: the assistant takes the conversation back. A conversation explicitly handed back to the assistant stays active.

During a pause, nothing automatic is sent to the client on that conversation: not the reminder the day before, not the follow-up the day after, not the reminder about a proposal, not a reply the assistant had already written and that was waiting to be sent (a few seconds to a few minutes of “thinking”, or the 24-hour window reopening). These messages stay in the thread, marked “Not sent”, with the reason. The general switch turned off counts as a pause everywhere: no automatic message, no team decision written up for the client. A reminder, follow-up or proposal reminder that has stayed pending beyond its time slot (a reminder “for tomorrow” about a service that has already started) is never sent: “Not sent: its time slot had passed.” A human has taken over, and the assistant does not interrupt them. If the assistant resumes early enough (within six hours of the planned time of a reminder or a follow-up, or before the start of the service for a proposal reminder), the message is then sent normally.

What also stops the assistant, regardless of you:

• the month's messages are used up (see 2.11);
• the account is suspended for an unpaid invoice, or the subscription is cancelled (chapter 14).

In these cases, a badge next to the switch says so (“Monthly messages used up”, “Stopped by the subscription”) and messages wait for the team in the inbox. Safety lines (AI disclosure, emergency, angry client) are sent even when the quota is used up.

2.10 The assistant settings#

Where: Settings › Assistant. Who: owner and managers.

There is nothing to set for its way of writing: the assistant adapts to each client on its own. It replies in the client's language, uses informal address if the client does and formal address otherwise, matches the length and tone of its messages to theirs, and works out your line of business from your catalogue. Why: an assistant that adapts to each person sounds more natural than a tone imposed on everyone, and no text typed in can slip into its instructions (a door closed to manipulation).

The page keeps only two things.

Status

• Assistant on / Assistant paused: saved straight away.

Clients and availability

• New clients only: the assistant only talks to people who were unknown when WhatsApp was connected; your regular clients stay in your hands. You can always hand one of their conversations over to the assistant from the inbox.
• Holiday mode: the assistant gives longer, honest time frames (“I'll get back to you during the day”) without stopping replying, and replies a little more slowly.

Concierge services that signed up before 12 October 2026. The old page offered a type of business, a tone, informal or formal address and a reply length. These choices have gone: a concierge service that had chosen something other than the default values (for example “Very formal” or “Formal (vous)”) now has the assistant adapting to each client.

Elsewhere in the settings:

• Settings › General › Assistant’s first name: the first name it writes under.
• Settings › General › Your clients’ languages: the first language in the list is used for a new client until they have written (for example, a British number gets English, but a dialling code whose language you do not speak gets your main language until the client's first message).
• Settings › General › Team language: the language in which the assistant asks the team questions and writes the records.
• Settings › Automations: reminder the day before, follow-up the day after, reminders about proposals (chapter 13).

What cannot be set, and why. The way of writing (see above), the threshold above which a reply is held, the time before questions are escalated (30 minutes) and the automatic resumption after a pause (6 hours) are set by Vetexis. A concierge should not have to adjust the mechanics of an assistant. Having no emoji is also a house rule.

2.11 What counts as a message#

Each plan includes a number of messages per month (1,000 on Essential, 5,000 on Pro, 15,000 on Network), counted from the first to the last day of the calendar month.

Counts as one message: each turn in which the assistant writes a reply itself: a reply sent, a reply held for approval (with its holding message), or a holding message when the reply was discarded. Three grouped bubbles count as a single message.

Do not count: silences (“ok thanks”), safety lines (AI disclosure, emergency, angry client), the acknowledgement of a request for codes, the acknowledgement of a media file, the holding message that follows the client's agreement to a proposal, and the fallback in case of a failure on our side.

At 90% and then at 100% of the month, management is notified. Once the quota is reached, the assistant pauses and messages wait for the team, until a message pack is bought or the next month begins (chapter 14).

2.12 Real-life examples#

A table at the beach club. An Italian client writes at 23:00: “Questo posto! Si può avere un tavolo la sera?” with a screenshot. On Pro, the photo is described; the assistant understands the place, asks for which evening and how many people (a single question), then creates the request. The concierge receives, in English: “Request #48: VIP evening table at Arena Beach Club, Saturday 23:30, 6 people, for Ferraresi. Available?”

A discount. “Can you do something on the price for the week?” The assistant promises nothing, says it is looking into it, and the request goes to the team as a quote to prepare.

The gate code. “What's the gate code?” The assistant replies that it is looking into it, and the concierge receives a priority question. It does not know the codes and does not know how they are passed on: this is deliberate.

An emergency. “My son fell in the pool, he's not responding.” Within a second, the assistant replies to call 112, alerts all the managers and creates an urgent request “call back straight away”.