AI agent for Airbnb hosts, explained: what it is and how it differs from a chatbot

An AI agent for Airbnb hosts is software that reads guest messages from every channel, checks the reservation and your property's knowledge base, answers or takes an action, and hands the thread to you when it is unsure. A chatbot only says what its script contains. Hotel Technology News (July 2025): agents handle 70 to 90 percent of routine inquiries.

Below: a definition you can quote, the three sources that back it, a table of agent versus chatbot versus rule-based automation, what an agent needs before it can work, five questions for any vendor, and prices checked on 26 September 2026.

What is an AI agent for short-term rentals

An AI agent for short-term rentals is software built on a large language model that reads guest messages from every channel, looks up the property's knowledge base and reservation data, decides on its own whether to reply, take an action (such as creating a booking) or hand the thread to the host, and does so in the guest's language within seconds.

Three things in that sentence separate an agent from everything hosts had before: it picks the next step itself, it has access to data and tools, and it is allowed to back out. Each is confirmed by sources that define agents in general, not just for hospitality.

Anthropic (the essay "Building effective agents", 19 December 2024) splits the field in two. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths". Agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks". The same essay calls customer support a natural fit for agents, because a conversation needs outside information and actions, and it warns to add complexity "only when it demonstrably improves outcomes".

OpenAI (the guide "A practical guide to building agents") is shorter: "Agents are systems that independently accomplish tasks on your behalf." It also rules things out. Applications that use an LLM but do not let it control the workflow, "think simple chatbots, single-turn LLMs, or sentiment classifiers", "are not agents". An agent, in OpenAI's version, has three parts: a model, tools and instructions.

The EU AI Act (Regulation 2024/1689, Article 3(1)) never uses the word agent, but its definition of an AI system describes the same thing: "a machine-based system that is designed to operate with varying levels of autonomy (...) infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions". A host who runs one is, in the Act's words, a "deployer" (Article 3(4)).

We do not quote Gartner: its article on AI agents refuses access without a login.

How an agent differs from a chatbot and from rule-based automation

The difference is who decides the next step. In a chatbot and in an automation, a person decided it while writing the script. In an agent, the model decides, based on the guest's message and the data it can reach.

A real example. A guest writes at 11:10 pm: "We're outside, the code doesn't work, the baby is asleep on my arm." The chatbot shows a menu with three buttons. The automation sends "Thanks for booking" the next morning. The agent looks up which unit the reservation is for, sends the code again with the lock instructions, and when the guest says it still fails, hands the thread to you with the whole context. Conduit counted, across 10 million messages, that roughly one in ten is about getting through a door, and in 45 percent of those the guest reports that something is not working.

Rule-based automation is not a bad choice. Booking.com says over 30 percent of all partner responses are already automatic replies, around half of manual responses use templates, and over 25 percent of partner-initiated messages are scheduled templates. It works as long as the guest asks exactly what you predicted. The agent starts where the template ends.

What an agent does in an ordinary week

It answers the same five topics that make up 54 percent of guest questions in Conduit's analysis: arrival and departure logistics (17.3 percent), booking and availability (10.3), property and area information (9.0), access and entry (8.6), verification and registration (8.5). Conduit looked at roughly 10 million messages from 68,625 properties in 78 countries between August 2025 and July 2026 and published on 3 August 2026.

From the same analysis: 24.3 percent of messages land on check-in day. 27.4 percent of guests who asked for the door code had already been sent it. 20.8 percent asked for a check-in time they had in the confirmation, 18.4 percent for a WiFi password they also had. About 11 percent of what reaches the inbox is not written by a guest at all. The agent takes exactly that part: the code, the time and the password, politely, for the third time, at 10:40 pm.

How much it handles alone depends on what it knows and what it is allowed to do. Hotel Technology News (21 July 2025) describes agents "capable of handling 70 to 90 percent of routine inquiries without human intervention". Aeve, which sells an agent to property managers, says on its home page that operators "typically reach 70 to 90 percent of guest enquiries on autopilot"; that is a vendor claim, not a study. ChatLab says its hotel customers "automate around 70 percent of routine questions", also its own figure. A sample of 434 conversations handled by one agent in September 2026 (Supalobby) came out lower: 46 percent of text messages were answered without the host, 25 percent ended as a draft to send, 30 percent went to the host. The gap between 46 and 70 to 90 has two causes: how much the agent was allowed to do alone, and how many facts about the property it was given.

Hostaway's 2026 report says 61 percent of short-term rental operators used AI in 2025, with guest communication first. Hostfully's study from May 2026 says nearly half of operators used AI in 2025, and 30 percent of property managers named AI guest communication as the technology they most want to explore in 2026.

What an agent needs before it can work

Four things: a knowledge base for the property, reservation data, access to the channels, and a rule for handing things to you. Take any one away and the agent turns into a chatbot, only a more expensive one.

1. A knowledge base. Directions from the station and the motorway, parking, codes and how the lock works, WiFi, check-in and check-out times, house rules, deposit, pets, bins, heating, where to eat and where the bus leaves from. The right standard is the one Tracify, a Polish vendor, states for its assistant: it "does not guess when something is not in the knowledge base". An agent without the fact should stay quiet and ask you, not improvise.
2. Reservation data. Who is writing, which unit, which dates, what they paid, what is outstanding, whether the unit is free the night before. Without it the agent cannot answer "can we arrive at 1 pm". It gets the data from a PMS or from the platform's booking notifications.
3. Channel access. Booking.com delivers guest messages and accepts replies through its own relay, so an agent can answer in the thread without a PMS. WhatsApp works through the Meta Cloud API, e-mail through IMAP and SMTP, Instagram through Meta's API. For Airbnb, ask the vendor directly: does it read the thread, read and draft, or read and send. The answer often differs by channel.
4. A hand-over rule. OpenAI names two triggers for passing control to a human: exceeding a failure threshold (for example, not understanding the intent after several attempts) and high-risk actions, "canceling user orders, authorizing large refunds, or making payments". In hosting that means refunds, discounts, complaints, damage and anything about safety. The agent must be able to say "let me check and come back to you" and actually hand the thread over with context, not just ping you.

A fifth thing people forget: language. Booking.com translates guest messages in the Extranet and the Pulse app into nine languages (English, German, French, Spanish, Dutch, Polish, Russian, Arabic, Portuguese), but only for reading; you still write the reply yourself. An agent replies in the language the guest wrote in.

How to tell whether a vendor sells an agent or a chatbot

Five questions where a yes or no tells you everything:

1. Can I paste my own house rules and directions, and does a question outside that knowledge get "I don't know, I'll ask the host"?
2. Does it see my reservations (dates, unit, price), and where does it get them?
3. On which channels does it send by itself, on which does it prepare a draft, and which does it only read?
4. What happens when a guest asks for a refund or writes that the place is dirty: who gets that thread and how fast?
5. What share of messages at existing customers ends without a human, and is that measured or from a brochure?

If question 1 gets "we configure scenarios", it is a chatbot. If question 4 gets "the agent resolves it", walk away.

Why speed is the whole point on Airbnb

Because Airbnb measures it. Your response rate is "the percentage of new inquiries and reservation requests you responded to within 24 hours in the past 30 days", and your response time is the average across all new messages in the same period. Superhost status requires responding to 90 percent of new messages within 24 hours, alongside a 4.8 rating and a cancellation rate under 1 percent.

Twenty-four hours is the rule; the guest's expectation is minutes. A drafted reply sitting in your inbox while you drive is worth less than a sent one, which is why the choice between drafts and autopilot matters more than the choice of vendor.

Where Supalobby fits

Supalobby is one example of the category: an AI agent plus one inbox for hosts. It reads messages from Booking.com, WhatsApp, Instagram, e-mail, Slowhop and Aloha Camp and replies in them by itself; on Airbnb it reads the thread and prepares the reply, which you send with one click. Reservation data comes from Beddesk or HotRes, and in Autopilot it creates bookings in them. It keeps a knowledge base per property and has a confidence gate: when unsure, it asks you and hands the thread over. In a sample of 434 conversations in September 2026 the median time to reply was 3.6 seconds. It has no calendar, no rates and no channel manager, and no IdoBooking, KWHotel or Smoobu integration today. Copilot 25 zł, Autopilot 49 zł net per unit per month (about 6 and 11 euro), 14-day trial, no card.

If you keep one sentence from this page, keep this one: an agent is software that picks the next step itself, and knows when not to.