Mr.GG as a conversation: the analysis Contents

Hostify

Mr.GG as a conversation: the analysis

What can happen in a conversation with an AI agent, what cannot, and why: channels, law, the 98 moments of a stay. A discussion paper; nothing here is decided.

November 2026

Laid out for A4 as well: Print, "Save as PDF", paper A4, background graphics on.

1 About this paper

Status: discussion paper, 2026-11-16. Nothing here is decided. The chapter "Ideas from the discussion of 2026-11-16" records the ideas of the discussion held the same day; the CTO called them "only ideas". The concept that grew from them is published at https://hostify-mrgg-concept.pages.dev.

The question: can the whole guest experience of Mr.GG (registration, documents, signatures, deposit, extras, entry instructions, the guide, check-out) be driven by the AI assistant in a conversation, instead of by a checklist of forms?

Sources: the interaction inventory in part-3-guest-experience.md (98 moments); annex-ai-agent-touchpoints.md; the research file wiki/raw/research/2026-11-16-conversational-check-in-precedents.md. Most facts in that research come from search excerpts, and the legal points are desk research. Each legal point below needs a lawyer per country before it becomes a product decision.

2 Ideas from the discussion of 2026-11-16

Record: wiki/decisions/2026-11-16-mrgg-chat-first-with-widgets.md (status Proposed).

The CTO's idea goes further than the target proposed in the chapter "The short answer". The product would be a chat with an AI agent that puts specialised widgets into the chat when needed, for the guest and for the manager alike. The manager does not fill templates: the manager talks with the agent about the property and photographs things (the remote control of the air conditioner, the coffee machine), and the agent builds a wiki of the property. Map and review services are used to build a guide tied to the property.

Three layers hold it together:

LayerWhat it isWho writes it
KnowledgeThe property wiki: free text, a source and a date on every fact, inherited from account to building to listingThe agent, from conversation, photographs, manuals, reviews and past guest questions
PolicyDeposit, accepted documents, compliance scheme, agreement textSet in conversation, stored as structured data, shown back for confirmation
StateWho registered, what is signed and paidThe system, never the model's reading of a thread

Door codes are not in the wiki the model reads; a rule releases them.

Four design questions were answered by Claude when the CTO delegated them; they are proposals:

  1. Core, chat and widgets only. The guest hub and the template editors are not rebuilt. Two views remain beside the chat: the guest's progress widget, which lists the steps and opens each step's widget directly, and the manager's sheet of policy and knowledge, where every stored value is visible with its source and can be corrected directly. Existing templates are converted automatically.
  2. Confirmation by risk. A fact about using something in the home, or about the area, is live at once, marked unconfirmed, and retired at the first contradiction. A fact about safety, access, money or rules is not stated to a guest until the manager has confirmed it. The manager gets a digest, not a prompt per fact.
  3. Quick buttons in the same chat are enough. No second surface; every step can be completed from the progress widget without typing.
  4. One agent runtime and one knowledge, three roles (manager, guest, cleaner), each with its own tools, instructions and data scope, enforced by the server.

The widgets: document capture, signature, payment, summary for confirmation, access card, photo upload.

What this would change in the specification, if it were adopted: parts 1 and 2 would stand as the deterministic core; part 3 would stop being a list of screens to build and become the content of the widgets and of the agent's tools; in part 4 the template editors and assignment screens would not be built, and its data and rules would become the policy layer. None of this is adopted; the four parts stand as written.

3 The short answer

Yes for the guest's experience. No for the literal reading "everything happens as chat messages".

  • Every moment where the guest learns something, chooses something or states a simple fact can happen in the conversation, on every channel.
  • Four kinds of moment cannot be chat messages, for reasons that are law and platform policy and will not change with better models: capturing an identity document, giving a signature the law requires, paying, and receiving an access secret. The first three need a small secure surface that the conversation opens and closes. The fourth can arrive in the thread but must be released by a rule, never by the model.
  • No vendor found does the whole check-in in the thread. All of them (Duve, Enso, Akia, Canary, Chekin, Conduit, Mews) use chat to invite and remind, then send the guest to a web view for everything else. Being conversation-led for the whole journey would be ahead of the market.

The target this paper proposes is conversation-led, card-assisted: one assistant that knows the state of the check-in, says what is needed next, takes what can be said, and opens a card for what cannot. The same card is a component in our own chat, a Flow in WhatsApp, and a small web view opened from an OTA thread.

4 Three meanings of "conversational"

LevelWhat it isWhat the guest feelsVerdict
A. Assistant beside the formsToday's hub stays the product. The assistant knows the check-in state, answers "what is left?", and sends the link to the right stepA helpful chat next to a formCheap, useful, and the necessary first step
B. Conversation-led, card-assistedThe conversation is the product. The assistant leads; structured and sensitive steps appear as cards inside or beside the thread, prefilled, and return to the threadOne conversation from booking to check-outThe proposed target
C. Pure threadEverything is a message: the passport photo, "I agree", the card numberA chat and nothing elseNot permitted: regulators, platform policies and card rules each forbid a part of it

5 What decides it: the channels and the law

5.1 By channel

ChannelWhat the thread can carryWhat must leave the threadNotes
Our own web chat (in the hub, on the booking site)Everything, as inline components we controlNothingThe only place where level B is fully "in the conversation"
WhatsAppText, buttons, lists, media; Flows: multi-screen forms in the thread with a camera picker whose files go encrypted to our endpoint, not into the visible threadPayment (a link; native payments exist only in India and Brazil)Policy forbids asking for card numbers and "other sensitive identifiers" in messages. Service messages inside the 24-hour window are chargeable since 2026-10-01. Opt-in is required
Airbnb threadPlain text and imagesEverything structuredPolicy for bookings from 2026-11-20: a host may not collect contact details, ID documents, financial or biometric data, except where a law requires it and the house rules disclose it; separate rental agreements are limited to listed cases; most hosts may not charge deposits. Links that take people off the platform are a violation outside those cases
Booking.com messagesText, PNG and JPEGEverything structuredLinks are stripped unless allow-listed per property. Since 2026-09-28 the guest's phone number is no longer passed on, which removes the bridge to WhatsApp or SMS
VrboText; links after confirmationEverything structuredNo rule on check-in tools was found
SMS, RCSText and a link (RCS adds buttons and cards)Everything structuredNot end-to-end encrypted
VoiceSpeechDocuments, signatures; payment by link or keypadThe same tools as the chat; the voice concierge proposal (P6)

5.2 By step

StepWhy it cannot be a plain messageWhat it becomes
Identity documentThe Italian data protection authority (2026-04-29) names document photos sent through messaging apps as a risk practice and requires deletion right after transmission. The Spanish authority has fined hotels for copying identity documents. A photo in a thread stays on the guest's phone, in backups and at the platform, and we cannot delete itA capture card that sends the image to our endpoint and keeps nothing in the thread
Identification in ItalyThe law requires real-time identification of the person against the document; remote is allowed only by live videoA live video step, or in person. Not solvable by any form or chat
Signature the law requiresSpain: everyone over 14 signs the traveller record. Portugal: "filled in and signed personally". Whether a typed "I agree" satisfies either was not foundA drawn signature on a card, until a lawyer says otherwise
Rental agreementA typed assent is a valid simple electronic signature in principle and the weakest evidence. It needs the exact text, its hash, the identity binding and a copy delivered to the guestAssent in the thread is defensible if the evidence is kept; a card is safer. A choice per manager
PaymentCard data typed in a thread pulls the channel into card-security scope; strong customer authentication applies to payment linksA hosted payment card or link. Once a card is saved with authentication, later extras can be charged on a "yes" in the thread
Door codeA model can be argued withA deterministic release: a service checks the state and sends the code into the channel bound to the reservation. The model can ask for the release; it never holds the code

One rule applies to the whole design whatever the channel: since 2026-08-02, Article 50(1) of the EU AI Act requires that the guest is told, at the first interaction, that they are talking to an AI. In an OTA thread the guest believes they are writing to the host, so it is not "obvious". Today's guest agent is instructed never to confirm or deny that it is automated (annex, prohibitions, line 101). That instruction and the law point in opposite directions.

6 The 98 moments, sorted

From the interaction inventory. Ids refer to the appendix of part 3.

Kind of momentExamplesIdsIn a conversation
Learn a fact about the stay or the homeDates, address, what is left to do, deposit terms, house manual, places nearby, check-out tasks, statusesI02-I04, I08-I09, I36, I41, I47, I50-I51, I54-I55, I59, I65-I68, I70, I72-I73, I76-I77, I79-I80, I84-I85, I87-I88, I92, I97-I98Said. This is what the assistant already does well, and better than tiles: the guest asks instead of searching
Make a simple choice or give one valueArrival time, which guest, document route, deposit or waiver, tasks done, ratings, languageI06-I07, I10-I14, I16, I19-I20, I44, I48, I56-I57, I69, I74, I81-I83, I86, I89-I91, I95Said or tapped. "We land at nine" is easier than a time picker
Give personal detailsNames, date of birth, nationality, address, email, phone, purpose of visit, the manager's own questions, invoice detailsI23-I28, I33-I34, I37-I38, I52-I53Said, then confirmed on a card. The model turns "two adults and our six-year-old, we live in Lyon" into draft fields. The guest confirms a summary. On Airbnb, only what a law requires
Give identity-document dataScan or photos, document number, issue and expiry, support number, visa, personal numberI15, I17-I18, I21-I22, I29-I32, I35Card only. The reading of the document fills the fields; the guest confirms. Never as a message
SignRead and sign the declaration; read and sign the agreement; get the signed copyI39-I40, I43, I45-I46Card, or assent in the thread for the agreement where the manager accepts the weaker evidence
PayDeposit or waiver; an extraI49, I71Hosted payment card. Extras by "yes" once a card is on file
Receive an access secretEntry instructions, door code, lock link, Wi-Fi password; trigger a lock actionI60-I64Delivered in the thread by the release rule, never composed by the model
Account and device housekeepingOpen the link, accept terms, sign in, saved stays, embedded-browser gateI01, I05, I58, I93-I94, I96Mostly disappears: the thread is the session

The count that matters: under the Spanish scheme a typical adult fills 16 to 19 fields, and a family of five would be asked about 85 questions if the conversation asked them one by one. That is the case against level C even if it were allowed. The conversation has to collect by reading and confirming: the booking already knows names and counts, the document reading knows most of the rest, and the guest corrects a summary instead of answering questions.

7 How it would be built

7.1 One state machine, three renderers

The check-in already is a state machine, though the legacy code never names it: places, requirement states, the ordered readiness gates, the unlock rule. The rebuild should make it explicit and give it one function: "for this reservation and this person, what is the next thing needed, and why?" The hub's checklist is one rendering of the answer. The assistant's next message is another.

Each step of the checklist is then built once, as a self-contained card with its own short-lived signed entry:

  • in our web chat it renders inline;
  • in WhatsApp it is a Flow;
  • in an OTA thread it is a small web view opened by an allow-listed link, which returns the guest to the thread.

This is the one recommendation in this paper that affects the rebuild now: if the checklist tasks of part 3 are built as embeddable, deep-linkable steps from the start, the conversational product later costs a renderer, not a rewrite.

7.2 What the assistant is given

Today the agent sees Mr.GG only through the PMS: the texts of the content templates, whether a check-in form is active, and the link. It does not see whether entry instructions are open, who has registered, whether the agreement is signed, what was bought, or whether the check-in was submitted. To lead a check-in it needs these tools, all scoped to the one reservation of the thread:

ToolDoesGuard
Read the check-in stateThe next required step and the reason; who has registered; what is signed and paidRead-only
Draft guest detailsWrites what the guest said into an unfinished place, as a draftNever completes a place; the guest confirms on the card
Set stay timesArrival and departureThe guest's own statement, recorded as such
Open a cardReturns the card or link for a named step: document, declaration, agreement, depositSigned, short-lived, bound to the reservation
Offer and sell an extraOffers from the catalogue; charges a saved card after an explicit yesPrice from the catalogue; a spend ceiling; "on request" items go to the manager
Request entry instructionsAsks the release service to send themThe service decides from state and time; the model never sees the code
Hand over to a personAs todayAlways available; the platform policies require it

7.3 Who is in the conversation

A reservation has several travellers and the thread has one person: the booker. The legacy product already answers this: an unfinished place can be filled by whoever holds the link, so the booker fills for the family. Signatures are the exception. Under the Spanish scheme everyone over 14 signs their own declaration, so each of them needs their own card, which the booker forwards. The conversation cannot remove that step; it can make it one tap per person.

7.4 Identity

The link is the permission today. In a conversation, the thread of that booking takes its place: an Airbnb or Booking.com thread is bound to the reservation by the platform, a WhatsApp number by verification. A new number that says "I am also the guest" is not bound. Access secrets go only to a bound channel.

7.5 The assistant starts the conversation

A checklist waits for the guest. A conversation can begin: after the booking, "to prepare your arrival I need three things", and then nudges that replace today's reminder schedule. Today's agent is forbidden to bring up the check-in form on its own initiative (annex, line 62). That rule was written for an assistant that answers; a conversation-led product reverses it on purpose, per manager.

8 A path in four steps

StepWhat shipsDepends on
0. AwareThe assistant reads the check-in state, answers "what do I still need to do?", sends the link to the exact next step, and tells the guest it is an AI. No new surfaceThe state function and deep-linkable steps in the rebuild; lifting two prohibitions
1. Conversation in our own surfaceInside the hub and on the booking site the conversation is the main view; steps render as inline cards; details are collected by reading and confirmingStep 0; the cards
2. WhatsAppThe same conversation with Flows for the structured steps; payment by linkStep 1; a WhatsApp business setup per manager; opt-in; the message costs
3. OTA threads and voiceThe assistant leads in the Airbnb and Booking.com thread; one allow-listed mini view for documents, signatures and money, on Airbnb only where a law requires them. Voice uses the same toolsStep 1; an answer from each platform's partner manager

Step 0 is small and worth doing whatever is decided about the rest.

9 Risks

  • The assistant's word binds the operator. A tribunal held an airline liable for what its chatbot said. An assistant that says "your check-in is complete" or "late check-out is fine" has said it for the manager. Statements of state must come from the state, never from the model's reading of the thread.
  • Argued releases. No real incident of a guest-messaging bot leaking a door code was found, and training exercises model exactly that failure. The release rule outside the model is the control.
  • Airbnb's policy of 2026-11-20 reaches Mr.GG itself, with or without AI: document collection, separate agreements and deposits are each limited. A legally required registration qualifies as an exception; a discretionary agreement or deposit for an Airbnb booking may not. This needs an answer from the partner manager for the product as it is today.
  • Wrong extraction. A model that fills a police registration from free text can fill it wrongly. The confirm card and the manager's view of what was read against what was typed are the controls. A registration is the guest's declaration, so the guest must see and confirm exactly what is sent.
  • Cost. WhatsApp conversations and model calls have a price per guest; a checklist has none. The measure is completed check-ins before arrival and the staff time not spent chasing them.
  • Evidence. The research found no controlled comparison of check-in by conversation against check-in by form. The best number for an in-thread form against a web page is one self-reported case. We would be measuring this ourselves; the simulation gate and the quality score proposed in the market review are how.

10 What this asks of the specification

Four things, each small if done at the start of the rebuild and expensive later.

  1. An explicit check-in state with one "what is next and why" function.
  2. Every checklist task as a self-contained step with a signed, short-lived entry of its own.
  3. Entry instructions released by a service that takes the reservation and the channel and decides alone.
  4. One owner for each of the four reservation fields the guest surface and the agent both write, with the writer recorded.

11 Questions for the discussion

None is settled. The chapter "Ideas from the discussion of 2026-11-16" proposes answers to questions 1, 5, 8 and 12.

  1. The target. Is level B (conversation-led, card-assisted) what "fully conversational" should mean for us?
  2. The first channel. Our own web chat inside the hub, where everything can be inline, or WhatsApp, where the guests already are?
  3. Initiative. May the assistant open the check-in conversation and nudge, per manager? Today it may not mention the check-in unless asked.
  4. Disclosure. The agent is told never to say it is an AI; the law has required the opposite since 2026-08-02. How is it worded, and who confirms the legal reading?
  5. Whose assistant is it? The same Auto Pilot each manager switches on and instructs, or an assistant that is part of Mr.GG for every reservation, with the manager's instructions on top?
  6. Signatures. Is a typed assent acceptable for the rental agreement, as a setting per manager? For declarations we assume a drawn signature until a lawyer says otherwise: agreed?
  7. Money on a "yes". May the assistant charge a saved card for an extra after an explicit yes in the thread? Up to what amount without a person?
  8. Reading and confirming. Do we accept that the model drafts registration fields from what the guest says and from the document, with the guest confirming a summary?
  9. The four shared fields. Who owns planned arrival, planned departure and the two marks: the guest's statement, the assistant's inference, or the manager?
  10. Airbnb. Who asks the partner manager how the 2026-11-20 policy treats our check-in page, agreements and deposits?
  11. Italy. Real-time identification is the law. Do we offer a live video step, or leave Italy to in-person check-in?
  12. The cost now. Do we accept building the checklist tasks as embeddable steps in the first version of the rebuild, so the later steps are renderers?
  13. Measurement. What is the number that says the conversation is better than the checklist: completion before arrival, time to complete, or staff time?