"Agentic AI" has become the phrase of the year in short-term rental software. Guesty describes its new features as steps towards an agentic layer. Independent vendors call their products AI agents. Most of the marketing stops there, without saying what an agent actually does differently when a guest writes in at 11pm.
This guide breaks it down. It follows one guest request through the five steps an agentic system runs, shows where each step usually fails, and explains why the right target is roughly three-quarters of requests on autopilot, not all of them. It applies whether you run 20 properties or 2,000, on Guesty, Hostaway or another PMS.
Chatbot, copilot, agent
Three labels get used interchangeably. They describe different amounts of work.
| What it does | Who finishes the job | |
|---|---|---|
| Chatbot or auto-reply | Answers a question from your listing or knowledge base | Your team, for anything that needs action |
| Copilot or drafts | Writes a suggested reply for a person to approve | Your team, every time |
| Agent | Works out what's needed, checks your rules, acts across your systems and closes the loop | The agent for routine requests, your team for the rest |
The difference shows on any request that needs more than words. A guest asks to arrive at 1pm. A chatbot quotes the 3pm check-in time. A copilot drafts "let me check with housekeeping". An agent checks the turnover, asks the housekeeper, waits for the answer, confirms with the guest and updates the reservation, without anyone on your team opening the conversation.
That's what people mean by autopilot for short-term rentals. "Autopilot" is the outcome, "agentic" is how it's achieved.
The five steps behind every request
Every request an agent handles runs through the same loop: context, classify, decide, act, learn. Most of the hard work, and most of the failures, sit in the first three.
1. Context: what's actually going on
Before anything else, the agent needs to know where this conversation stands. Is the guest waiting on something from you? Did someone promise to come back to them yesterday? Did they ask two things last night and only get one answered? Which property, which booking, how many days until arrival?
An agent that reads only the newest message answers the wrong question politely. "Thanks, see you Friday!" looks like a pleasantry. If the guest asked about a cot two messages earlier and nobody replied, it's a reminder.
2. Classify: which policy applies
Next the agent decides what kind of request this is, and so which of your policies governs it. A mature playbook has dozens of policies grouped into a handful of areas (arrival and access, the stay, problems and maintenance, money and changes, and so on). Each one says three things: what the agent can do alone, what it must check first, and what always goes to a person.
This is where general-purpose AI trips up. "Can we get in a bit earlier?" and "Is there somewhere to park when we arrive?" look alike to a general model: both are about arrival. To an operations team they're completely different. One depends on the cleaning schedule and needs a housekeeper's yes. The other is a fact from the listing. Get the classification wrong and every step after it goes wrong too.
3. Decide: reply, check first, or hand over
With the context and the policy in hand, the agent makes one of three calls:
- Reply. The answer is known and safe to give. Wifi, parking, directions, house rules.
- Check first. The answer depends on someone or something else. Early check-in depends on the turnover. A cot depends on whether one is free. The agent asks, waits and then replies.
- Hand over. The request needs judgement. Refunds, complaints, a guest who is upset or unwell, anything outside your rules. The agent passes it to the person responsible for that property, with a summary of what's been said and what the guest wants.
At portfolio scale this decision happens thousands of times a day, and it's the one that decides whether you can stop checking the agent's work.
4. Act: do the work, not just the reply
Then the agent carries it out. It sends the reply, messages the housekeeper on WhatsApp, creates the maintenance task with the address and access details, updates the arrival time in Guesty or Hostaway, and tells the guest when it's done.
5. Learn: so the questions get fewer
Finally, anything confirmed along the way is kept, per property and per rule. The housekeeper said cots are in the hall cupboard at Harbour House, so the next guest who asks gets an answer straight away. A manager corrected a reply, so the correction applies from now on. Over time the agent needs to ask less, and your team hears from it less. We've covered what self-learning should mean, its risks and how to test it.
Why 70 to 80% is the sweet spot
It's tempting to judge an agent by how much it automates. The better question is how it behaves on the requests it shouldn't touch.
One wrong reply in a hundred (a late checkout agreed when the turnaround is tight, a maintenance visit promised that nobody booked) costs more than the 99 right replies saved. The guest is let down, and housekeeping or maintenance end up dealing with a promise they never knew had been made. The volume makes this concrete. A property typically brings 10 to 20 guest messages a week.
| Portfolio | Guest messages a week | One in a hundred |
|---|---|---|
| 50 properties | 500 to 1,000 | 5 to 10 a week |
| 300 properties | 3,000 to 6,000 | 30 to 60 a week |
| 1,000 properties | 10,000 to 20,000 | 100 to 200 a week |
At scale, a 1% error rate means hundreds of bad promises a month. That's why the aim isn't the highest automation rate possible, it's the highest rate that never costs you a guest.
A capable agent can handle almost any guest request accurately. That doesn't mean it should. Aiming for 70 to 80% on full autopilot, with refunds, complaints, frustration and anything outside your rules going to your team with the context gathered, is the balance that holds up across portfolios. Escalating correctly matters more than replying. A figure close to 100% should make you ask what's being answered that shouldn't be. Our guide to measuring automation rate sets out how to count it honestly.
Where agents work across STR operations
Guest messaging is where agents started, and still where they save the most time. The same loop now applies across the wider operation.
- Guest communication. Questions, problems and requests, before, during and after the stay, in any language.
- Coordination with your team and suppliers. Asking a housekeeper or handyman, waiting for an answer, confirming back to the guest. See automating the tasks behind every stay.
- Staff scheduling. Planning each person's day from your staff list, and re-planning it when someone calls in sick. See AI staff scheduling for short-term rentals.
- Guest experience. Reading every stay's messages and reviews, scoring what went wrong and pointing at the one fix that matters most. See scoring every stay.
- Revenue. Offering early check-in, late checkout and gap nights when the calendar allows. See AI upselling.
What stays with people
Agentic doesn't mean unattended. Someone still owns each property, sets the policies and makes the calls the agent brings to them. The pitch isn't fewer people. It's the same people spending their day on the cancelled flight, the upset family and the owner conversation, rather than on door codes.
Tone matters here too. A reply can be correct and still cost you the guest. Service is getting them the wifi code. Hospitality is how it felt to ask. An agent should sound like your team, learned from how your team already writes. We cover this in why AI guest messaging goes wrong.
How to evaluate an agentic tool
Ask any vendor to walk one request through all five steps, live, on one of your properties. A good one to use: a guest who asks to arrive early on a day with a tight turnover. Then ask:
- Context. Show me a conversation where the guest was still waiting on something. Did the agent notice?
- Classification. Show me your policies. What does the agent do alone, check first and hand over for each?
- Decision. Show me ten conversations it handed to a person, and why.
- Action. What did it change in the PMS? Who did it message, and how?
- Learning. What did it learn this week, where from, and how would I remove something?
Then run it on a sample of your own requests before you commit. Our buyer's guide has the full list of questions, and the market comparison shows which tools are built this way. For large portfolios, see rolling out an AI autopilot across hundreds of properties.
Sources
- Guesty, ReplyAI Autopilot and AI Task Creation launch, describing an "agentic AI layer" (April 2026): press release
- Will Guidara, Unreasonable Hospitality (2022), on the difference between service and hospitality