AI agent for Airbnb hosts: a Hermes field story

The starting point: simple messages, scattered everywhere

The owner was not struggling because he did not know how to manage his properties.

He already had good operating habits. Property instructions. A calendar. A cleaning agency. Standard replies. Rules for check-in, late checkout, unusual requests, small incidents and cases that required approval.

The issue was that this knowledge lived in too many places.

One guest wrote on Airbnb to ask if they could drop off luggage before check-in. Another sent an email about parking. A third used WhatsApp to say they would arrive late. The cleaning agency needed to know that a baby bed had to be installed. The calendar had to confirm there was enough time between two stays.

Each time, the owner had to run the same mental loop:

• find the booking,
• identify the property,
• check the dates,
• reread the instructions,
• decide whether the request was inside or outside the protocol,
• reply to the guest,
• notify the cleaning team if needed,
• remember what had to be followed up.

The load does not come from one message. It comes from the fact that every message forces the owner to reopen the whole context.

That is exactly the type of problem described on my AI agents for Airbnb hosts and short-term rentals page. The goal is not to add yet another interface. The goal is to connect guest messages, calendars, rules, cleaning coordination and owner decisions into one controlled system.

What had to be scoped before connecting Hermes

The tempting shortcut would have been: "Hermes will read messages and reply."

That would have been too fast.

An agent that replies without a frame can create more problems than it solves. In short-term rentals, a bad answer can promise an impossible arrival time, miss a cleaning constraint, accept an exception the owner usually refuses, or mishandle a sensitive case.

So we started by writing down the real operating protocol.

For each property, we had to clarify:

• check-in and checkout times,
• possible flexibility,
• when luggage drop-off is accepted,
• conditions for late arrival,
• access instructions,
• available equipment,
• information to send to the cleaning team,
• requests that can be accepted directly,
• requests that must go back to the owner,
• message tone and wording.

This is not admin work around the project. It is the core of the project.

A Hermes agent is useful only if the owner rules are readable by the system. If the rule is only "I usually accept, except when I feel it will be complicated", the agent cannot act cleanly. It can prepare, but it should not decide.

This is close to the logic I described in my article on human approval for AI agents in production. You do not approve "AI" in general. You approve actions. Answering a question about wifi does not carry the same risk as accepting late checkout, managing a dispute or promising compensation.

What Hermes does when a message arrives

Once the frame is clear, Hermes can work.

When a message arrives, the agent first identifies it: channel, guest, property, booking, relevant dates, language, apparent urgency and request type.

Then it retrieves the context it needs. It checks the booking. It verifies Google Calendar. It finds the property instructions. It applies the owner rules. It checks whether an action is needed for the cleaning team or another provider.

Then it classifies the request.

Some requests are framed and can receive a direct answer. For example: repeat the address, send access instructions, confirm an already available item, share arrival information or answer a frequent question.

Other requests should be prepared, but not sent without approval. For example: accepting a very early arrival, approving late checkout when the cleaning window is tight, responding to a complaint, handling a commercial request, or dealing with an incident.

Other cases should not be automated. In those situations, Hermes groups the facts and escalates the case to the owner.

The important point is not that the agent "can reply". The important point is that it knows when to reply, when to prepare, and when to stop.

That is the difference between a chatbot and a business agent.

A chatbot answers in a conversation. A business agent moves a process forward. Here, the process crosses several tools: messaging, calendar, internal rules, cleaning coordination, owner follow-up and daily summary.

Cleaning is not a detail, it is part of the product

In short-term rentals, cleaning is often treated as a separate operation.

In practice, it is one of the places where automation has the most value.

A simple guest request can change field work. Baby bed. Extra linen. Late arrival. Early checkout. Luggage storage. Stain reported. Missing equipment. Window that does not close properly. Photo sent by the guest.

Before, the owner had to read the message, understand the operational impact, notify the cleaning agency, sometimes fill in a form on the agency website, then remember that an action was expected.

Hermes now handles this coordination when the frame is clear.

The agent prepares the useful information: date, time, property, booking name, instructions, options to provide, important details, possible photos or notes. When the cleaning agency tool needs to be filled in, it does that too according to the agreed protocol: property, date, time, instruction and useful data.

This changes a lot.

The guest does not need to know that their message triggered a cleaning action. The cleaning agency does not need to dig through conversations. The owner does not need to act as a human router between everyone.

This is where workflow automation meets the AI agent. AI understands the message and chooses the action. The workflow executes the action cleanly in the tools.

What the agent does alone, prepares or escalates

The project really took shape when we separated three autonomy levels.

First level: Hermes acts alone.

These are repetitive, low-risk requests already covered by the rules. A door code question. Standard arrival information. A house-rule reminder. Confirmation of available equipment. A standard cleaning handoff.

Second level: Hermes prepares.

These are requests where the agent can do 80% of the work, but the human keeps the final decision. It drafts the answer, checks the calendar, reviews the instructions, summarizes the risks and proposes an action. The owner receives a clear message with the information needed to approve quickly.

Third level: Hermes escalates.

These are sensitive cases: conflict, noise, damage, out-of-protocol request, refund, compensation, unhappy guest, calendar doubt, safety issue or commercial decision.

This separation avoids two common mistakes.

The first mistake would be to automate everything, then discover the problems later. Bad idea.

The second mistake would be to approve everything manually, until the agent becomes one more notification layer. Also a bad idea.

The useful zone sits between the two. Give autonomy where rules are stable. Keep human approval where the relationship, revenue or risk is involved.

This is the same logic as in my field story about the AI agent that creates quotes from Telegram. In both cases, the agent does not replace the responsible person. It prepares the work, asks the right questions and executes what is framed.

The daily summary: less noise, more control

One of the most useful deliverables is not an automatic reply.

It is the daily summary.

Every evening, the owner receives a simple recap:

• upcoming arrivals,
• departures to watch,
• important exchanges,
• completed tasks,
• messages awaiting approval,
• points to monitor,
• cleaning or maintenance items still open.

This summary matters because it changes the owner's posture.

Before, he had to stay inside the flow. Read messages, keep details in mind, check whether something had been forgotten, jump back into each channel.

With the summary, he manages by exception. He sees what deserves attention without rebuilding the whole story.

This is not only comfort. It is a condition for scaling.

When everything relies on the owner, each new property adds mental load: more messages, more cleaning, more coordination, more risk of forgetting something. When the protocol is clear and the agent holds the thread, adding a fifth or sixth property becomes much more realistic.

The ROI: time saved, mistakes avoided, capacity to grow

I will not turn this case into a magic promise.

The ROI does not come from an agent that "replies with AI". It comes from four concrete gains.

The first gain is time. Based on early feedback, we are already talking about several hours saved each week per property. Not because every answer takes an hour. Because small interruptions add up.

The second gain is fewer mistakes. A forgotten cleaning instruction, a rushed answer, an arrival accepted without checking the calendar, a sensitive message handled when tired. These are small mistakes, but they become expensive when they damage the guest experience or field coordination.

The third gain is response speed. In short-term rentals, a worried guest rarely sends only one message. If there is no answer, they follow up, switch channels, call, or start the stay with a poor impression. Answering quickly, when the reply is framed, improves the experience without requiring the owner to be available all the time.

The fourth gain is the ability to add properties. This may be the most important one. If each new property adds the same mental load, growth becomes painful. If part of the protocol is handled by the agent, growth remains demanding, but it does not multiply chaos at the same speed.

A good AI agent for Airbnb hosts is therefore not only a support tool. It is lightweight infrastructure for managing multi-property operations.

The guardrails that really matter

This kind of agent must stay carefully framed.

I mostly look at five points.

The first: access. The agent should read and act only where needed. Every connected channel must have a clear reason.

The second: logs. When Hermes replies, prepares, escalates or fills a tool, you need to understand what happened.

The third: approvals. Sensitive cases should come back with facts, not a simple "yes or no" button. The owner needs to see the request, context, applied rule, risk and proposed action.

The fourth: platform limits. Not every channel offers the same integration possibilities. When direct access is not reliable or allowed, the agent can prepare, centralize and request approval instead of forcing fragile automation.

The fifth: rule maintenance. An agent that is reliable today can become less reliable if instructions change and nobody updates them. Owner rules, cleaning instructions and exceptions must remain alive.

That is why I prefer presenting Hermes as a framed system, not as vague "autonomous AI". Autonomy is valuable only when it is readable.

Where to start if you manage several properties

If you manage several Airbnb properties, gites, furnished stays or seasonal rentals, I would not start by automating everything.

I would start with one flow.

The best first flow often has three traits:

• it happens every week,
• it requires checking several pieces of information,
• it creates a real cost when forgotten.

For some owners, that will be pre-arrival guest messages. For others, cleaning coordination. For others, the daily summary or out-of-protocol requests.

Then I would write the rules down. Not a perfect manual. Just enough structure for the agent to know what to do in 80% of cases, and what to escalate in the remaining 20%.

Then I would test on real cases before allowing more autonomy. Real messages are always messier than examples. Guests write quickly, change channels, forget details and mix several requests. That is normal. The agent must be built for that reality.

If the first flow holds, we can expand. Messages, calendar, cleaning, incidents, summaries, reviews, maintenance. Not as a large rebuild. As a progression around what already works.

Conclusion

This project confirms a simple idea: short-term rentals become hard to scale when the owner remains the only handoff point between guests, platforms, calendars and providers.

Hermes did not remove the owner's role. It moved that role to the right place.

The agent handles the repetitive work. It prepares decisions. It escalates sensitive cases. It keeps the thread between messages, calendars and cleaning. And every evening, it turns a day of notifications into a summary the owner can actually use.

If you want to explore this kind of system, the natural next step is the Airbnb hosts and short-term rentals page, then Hermes, AI Agents and Automation, depending on how much autonomy you want.