Most AI guest messaging tools look good in a demo. The trouble starts a few weeks after go-live, when real guests are writing in at 2am, half the portfolio is turning over on the same Saturday, and someone on your team is quietly reading every message before it goes out "just in case".
This guide covers the problems operators run into most often, why they happen, and how to check for them before they cost you a review. It's written for teams on Guesty or Hostaway, but none of it is specific to one PMS or one vendor.
The automation rate is easy to inflate
The number that matters is how much work gets done without a person. That's the time your team gets back, and it's what lets a portfolio grow without the support team growing with it.
It's also the number most often stretched. Three habits make it look bigger than it is.
Counting drafts. If the AI writes a reply and someone reads it and presses send, a person still did the work. Some tools count those as automated anyway.
Counting only the reply. A guest asks for an extra set of towels. The AI says "of course", and someone on your team still has to tell housekeeping. The message was answered, but the request wasn't handled. A fair count includes the work that follows: the task raised, the cleaner moved, the booking updated.
Choosing the denominator. A figure measured only on "eligible" categories, or only on properties where the AI is switched on, or only on the best customer, can't be compared with anything else.
Most systems manage a fairly small share of the real work, and the hardest requests, like early check-ins, are usually the ones sent back to your team. Those are exactly the ones you'd most want handled. We've written up how to measure automation rate properly, so you can apply the same yardstick to every tool, including the one in your PMS.
That said, the goal isn't 100%. Some conversations are better handled by a person, and a good system knows which.
Accuracy problems rotate rather than stop
An automation rate only counts if the answers are right. One wrong reply in a hundred is enough to stop a team trusting the system, and once that happens everyone goes back to checking every message. You're approving drafts again, and the saving is gone.
The pattern operators describe is consistent. They report a mistake, it gets fixed, and a different one appears the following week. Months in, the list is as long as it was at the start. That's because the cause usually sits in how the system reads messages, not in a missing fact.
A few message types cause most of the trouble:
- Two questions in one. "Can we check in early, and is it OK to bring the dog?" A system that matches one topic per message answers the first and ignores the second.
- A request that isn't phrased as one. "OK, see you at 12pm" looks like a pleasantry. If check-in is at 3pm, it's an early check-in request. Reply "Great, see you then!" and the guest turns up three hours early, expecting someone to greet them.
- The same request in different words. "We'll be there around noon" and "Can we drop our bags early?" need the same check against the turnover. Rules written for one phrasing miss the other.
- Context from earlier in the thread. The guest mentioned a pram two messages ago. The reply about the third-floor walk-up should notice.
- A guest still waiting on you. Yesterday the guest asked two things and only got one answered. Today they write "Thanks, see you Friday!" A system that reads only the newest message replies warmly and leaves the open question hanging.
- Requests that look alike but aren't. "Can we get in a bit earlier?" and "Is there parking when we arrive?" are both about arrival. To a general model they look similar. To your operation, one needs a housekeeper's yes and the other is a fact from the listing.
Most of these failures happen before the reply is written: in reading the context and deciding which policy applies. We walk through those steps in agentic AI for short-term rental operations.
The cost goes up once the system does more than reply. A late checkout agreed when the turnaround is tight, or a maintenance visit promised but never booked, lands on housekeeping or maintenance. They end up dealing with a promise they didn't know had been made. A cleaner sent to the wrong unit is a wasted visit and a scramble across two teams.
That isn't an argument for keeping coordination manual. It's an argument for being sure the system is accurate before you let it act on anything.
Be wary of being told the AI will improve if you give it a few more months. Every guest who stays during those months is on a real stay, and a reply that lands badly shows up in their review. It's worth asking what happens to guests during any training period, and whether you can see why the system replied the way it did. If it can show its reasoning and the source it used, a mistake can be traced to something you can fix that day.
Escalation is the part that earns trust
Buyers ask a lot about what an AI can answer. Fewer ask what it won't attempt. That second question is what decides whether you can stop watching it.
The failures that hurt are rarely wrong wifi passwords. They're the AI promising an early check-in housekeeping can't deliver, trying to settle a damage dispute on its own, or answering the literal question while missing that the guest is upset. The question is whether the system spots these before it replies, or whether you find out from the guest.
The arithmetic makes the point. A property brings 10 to 20 guest messages a week. Across 500 properties, one wrong reply in a hundred is 50 to 100 a week, each a promise someone else has to sort out. One of those costs more than the 99 right replies saved.
A system that automates 80% of messages sounds impressive. What you need is one that automates 70 to 80% and reliably hands over the rest. The automation rate tells you how much it does. The escalation logic tells you whether you can trust it.
The PMS vendors have settled on similar boundaries. Hospitable's auto-reply deliberately doesn't answer availability, early check-in, late checkout, date changes, refunds or payments. Hostaway's guidance treats early check-in and late checkout as medium risk to review, and refunds and compensation as high risk to handle personally. Those are sensible defaults, but they also show where simple auto-reply stops. Many of the requests guests care most about need a person, or a system that can check the calendar and the turnover before it answers.
Good escalation has a few traits:
- It's specific. The conversation goes to the person responsible for that property, not to a shared inbox nobody owns.
- It comes with context. A short summary, what the guest wants, and what's already been said, so nobody starts from scratch.
- It closes the loop. Once someone decides, the guest hears back and the PMS is updated. A request that sits in a queue after escalation is still unhandled.
- You can tune it. Some teams want every message about money in front of a person. Others are happy for the AI to handle standard fees and only flag exceptions. It's your call, and you should be able to change it.
While you're still building trust, it's better for a system to escalate too much than too little. You can loosen it once you've seen how it behaves.
Tone is the service, not a finishing touch
Will Guidara, who ran Eleven Madison Park, puts it neatly in Unreasonable Hospitality: service is black and white, hospitality is colour. In a guest inbox, service is getting the guest the wifi code. Hospitality is how it felt to ask.
Here's the same question answered both ways. A guest writes: "Hi! Landing at 1pm on Friday. What time can we get in, and how do we get the key?"
Service: Check-in is from 3pm. The door code will be sent on the day of arrival.
Hospitality: Hi Anna! Check-in's from 3pm, and I'll send the door code over on Friday morning so it's waiting for you when you land. If you're early, there's a lovely café two doors down to leave your bags at. Safe travels ☀️
Both are correct. Only one makes the guest feel looked after. Most AI replies manage the first and lose the second, and every operator's AI should sound like that operator, learned from how their own team already writes.
A reply can be factually right and still read as machine-written. Guests notice faster than most operators expect: long, polished paragraphs arriving seconds after a short question, the same phrasing for every guest, and enthusiasm that doesn't match the situation.
The cost is quiet. A guest who senses a bot rarely complains. They just don't leave the review, don't book again and don't recommend you. Communication scores drift from 4.9 to 4.8 to 4.7, and it's easy to blame the market.
Follow-through counts the same way. A guest told the cleaner would arrive at four, who then has to chase at five, has had a worse stay than the transcript suggests. For a guest, the message and what happens afterwards are the service. Both are your brand.
The test is simple. Read whole conversations from start to finish, the way a guest would, not the way an operator checks facts. Does it sound like your team? Does it pick up the small things a person would notice? Then check what happened next: was the job raised, the person told, the booking changed?
Voice agents aren't ready for most guests
AI phone agents are now on offer to the industry, and some demos are impressive. Our view is that the technology isn't yet at the point where most guests would choose it over a person. The short pauses, slightly wrong emphasis and the feeling of talking to a machine come through on a call in a way they don't in a message.
That's an opinion, and vendors who sell voice would disagree. If you want to try it, keep it to a narrow job such as out-of-hours triage, make it easy to reach a human, and measure bookings and complaints before and after. If those numbers move the wrong way, the saving on call handling isn't worth it.
Manipulation and promises it can't keep
Anything that reads guest messages and replies on its own can be steered by what guests write. In April 2026, 404 Media reported a guest who got a host's AI to reply with a French toast recipe during a dispute. It's funny until the same trick is used to get a refund agreed, a discount offered or a door code sent to the wrong person.
A few safeguards matter more than clever prompting:
- Guest text is information, never instructions. The system shouldn't change what it's allowed to do because a message says so.
- Commitments need permission. Refunds, discounts, date changes and anything involving money or access should need a policy that allows it, or a person.
- Disclosure. In the EU the AI Act expects people to be told when they're dealing with an AI. Elsewhere, being open about it is still the safer choice. Airbnb has said it allows approved third-party tools to reply for hosts, but a guest who feels misled is a guest who leaves a bad review.
The cost nobody puts on the slide
Volume adds up quickly. A property averages around ten guest messages a week. Across 1,000 units that's 40,000 or more a month, before any of the coordination messages that follow.
That shows up in two places. If you're building your own system, every message has to be processed, usually with the conversation, the reservation and the listing details attached. Unless the pipeline is carefully optimised, token costs climb fast. (Our build-your-own guide has a cost model.)
The bigger hidden cost is people. If someone has to keep reviewing outputs, updating rules and maintaining a knowledge base for each property, you've added headcount to run the tool that was meant to save it. Ask any vendor how much of that falls on you, and whether it grows with every property you add.
How to check it yourself
Whether you're evaluating a new tool or auditing the one you have, the method is the same. A week or two is usually enough. All the data you need is already in your PMS.
- Take a random sample. Around 200 guest requests from a normal period, not hand-picked easy ones.
- Count real automation. For each request, did it get done without anyone on your team stepping in? Include the follow-on work: the task, the cleaner, the booking change. Drafts someone approved count as manual.
- Log every wrong reply. Note the date, the message and what went wrong. Report each one and watch whether the list shrinks week by week. If new problems arrive as fast as old ones are fixed, that's your answer.
- Check what it escalated, and what it should have. Look for complaints, refunds and upset guests it handled alone.
- Read ten full conversations as a guest. Judge the tone and the follow-through, not just the facts.
- Count the upkeep. How many hours a week did your team spend correcting, reviewing or configuring it?
Put the real automation share alongside your team size and hourly cost, and you'll know what it's worth to you each month. If the numbers don't stack up during the trial, don't sign on the promise that they will later. For the questions to ask alongside this test, see how to choose an AI autopilot. For the coordination side, see automating the tasks behind every stay.
Sources
- Will Guidara, Unreasonable Hospitality (2022)
- 404 Media, Airbnb hosts don't want to talk to guests anymore, are outsourcing messages to AI, April 2026
- Hospitable, AI auto replies for guest messages
- Travolution, Hostaway launches AI auto-reply tool for vacation rental managers
- Airbnb, Superhost requirements and response rate