AISMITHAI Practice — Built World
AS-001Built work · Case

HostMate. Every guest, every language, every hour.

ClientDubai short-term rental operator
ScopeAI guest support system
Languages30+, auto-detected
Coverage24 / 7
StatusIn Service
Inputs — fragmented
  • Guest message, 02:14, Mandarin
  • Booking data, three platforms
  • Building access rules, PDF
  • House manual, out of date
  • Staff WhatsApp, overloaded
Output — structured
  • Answer in the guest's language, seconds
  • Grounded in verified property knowledge
  • Escalated to a human when it matters
  • Every conversation logged & reviewable

The problem

Short-term rental guest support is repetitive by nature — check-in codes, Wi-Fi passwords, how the stove or TV works, whether pool towels are provided, parking, house rules — the same handful of questions, asked hundreds of times a week, in whatever language the guest happens to speak. Handling that volume by hand doesn't scale by getting smarter, it scales by hiring: most operators plan for roughly one support staff member per 30–40 units. A guest landing at 2 a.m. speaking Mandarin is just the sharpest edge of a bigger problem — the same questions, asked constantly, at a volume that keeps forcing new hires.

What we built

HostMate is an AI guest-support layer that sits inside the operator's existing messaging channels. It detects the guest's language, answers from the operator's own knowledge — building access, check-in codes, house rules, local recommendations, payment and booking specifics — and hands off to a human the moment a conversation needs one. It doesn't pretend to be a person, and it doesn't guess: when it isn't certain, it escalates.

  • Grounded answers only. Every reply is drawn from the operator's verified property knowledge base — not from the model's imagination.
  • Human handoff by design. Complaints, refunds and edge cases route to staff with full conversational context, translated.
  • Owner-visible. Every conversation is logged, searchable and reviewable — support became an asset instead of a black box.
LIVE CONVERSATION · ANONYMISED 02:14 GST
Guest · RUЗдравствуйте! Мы приехали раньше, багаж очень тяжёлый. Можно оставить вещи где-то до заселения?
HostMate · RUЗдравствуйте! Конечно. В вашем здании есть комната хранения багажа на уровне P1 — консьерж примет вещи по номеру брони. Заселение сегодня возможно с 13:00 вместо 15:00, бесплатно. Подтвердить ранний заезд?
Guest · RUДа, пожалуйста!
HostMate · RUГотово — ранний заезд в 13:00 подтверждён. Код доступа придёт сюда в 12:30. Хорошего отдыха в Дубае!

It isn't a chatbot. It's a self-learning system.

Most "AI assistants" answer from a knowledge base someone has to write and maintain by hand. The moment reality changes — a new check-in procedure, a building's Wi-Fi password — the answers go stale until a human updates them. HostMate maintains its own knowledge base. A team of specialised agents watches the operation, learns from every conversation a human resolves, and rewrites the knowledge base under review — continuously, while the business runs.

We call this engine the AISMITH Learning Loop — self-learning, multi-agent systems that get sharper with use, not static chatbots with a real-estate skin. This is the hard part, and it is exactly what AISMITH builds.

AS-001 · THE AISMITH LEARNING LOOP 5–7 AGENTS · BY CONFIGURATION
01
ResponderAnswers the guest from the current, vectorised knowledge base — in the guest's own language, in seconds.
Live
02
WatcherMonitors every thread and locks focus the instant a human operator steps in — the signal that something new is being handled.
On handoff
03
ExtractorOnce the conversation is resolved, mines it for the genuinely reusable question–answer pairs — knowledge, not small talk — and writes them, formatted, into the Q&A ledger.
On resolve
04
CuratorA deeper model vets each candidate against embedded quality logic — is it grounded in what happened, general enough for the next guest, and not already known — before anything ships.
Quality gate
05
EmbedderApproved knowledge is vectorised into the retrieval base; the next incoming question is matched against it by meaning, not keywords.
Vectorise
Back to the Responder — now answering from what it just learned. Every resolved conversation makes the next answer sharper.
Built specifically for short-term rental, where the same twenty questions repeat daily but the details keep changing. Agent count scales with the deployment — language detection, escalation routing and more come online by configuration.
AS-001 · Q&A LEDGER — HUMAN REVIEW CAPTURED FROM THE LIVE ADMIN
HostMate Q&A moderation queue — the Curator review step, showing real guest questions, quality scores and approve/reject, guest identifiers blurred
The Curator step, in the operator's real admin: candidate Q&A mined from resolved conversations, each with a quality score and a human approve/reject before it enters the knowledge base. Guest identifiers are blurred — the data is real.

What changed for the operator

Guests served1,038
Q&A learned5,740
Properties37
Live run8 months
  • Every guest recognised and answered in their own language — the building, the booking and the guest's history resolved automatically from a single incoming message.
  • The knowledge base grew itself: 5,740 question–answer pairs mined and reviewed from real conversations, so answers got sharper the longer it ran.
  • Resolved 40–50% of guest questions without a human from day one, climbing to 80% within six months as the knowledge base learned from real conversations — only the emotional, financial or unusual cases went to a person.
  • Cuts the staffing ratio, not just the hours. A portfolio needing two operators at the standard 30–40-units-per-operator benchmark can typically run on one — same coverage, faster response, every language, at a fraction of the cost.
What didn't work in v1

The first version answered too much. It tried to handle complaints conversationally, and guests could feel it. We rebuilt the escalation logic so that anything emotional, financial or unusual goes to a human immediately — with the AI briefing the staff member instead of replacing them. The lesson is now practice policy: automate the routine, never the relationship.

Guest messages eating your nights?

Bring us one workflow →