In this piece
TWO MORE NIGHTS
Let's say a guest messages at eight at night asking to stay two more nights. The calendar shows both nights free, so the AI's reply says yes, and it's gone in seconds. It's polite, and everything it says about the flat is true. What it doesn't know is that your maintenance contact said on WhatsApp this afternoon that they're coming to replace the shower the day after this guest was due to leave. Your guest wakes up to a plumber at the door and a bathroom they can't use, your VA takes the complaint, and the review mentions it.
Where did that go wrong? It wasn't the writing. A new VA on their first evening shift would've sent the same reply, because they'd have read the same calendar. It was right about the property and wrong about the situation, and that's the risk with any message that goes out without your team in it.
Why writing the reply is the easy part
Most people think the risk with AI guest messaging is tone, or the AI making something up. Both happen, and you'll spot both the first afternoon you read its drafts. The risk that gets past you is that a guest message is rarely just a question. It's a job for someone, a change to a booking, money, or a problem that needs an owner, and the reply is the last step of all of that.
Sort a week of your own inbox by what each message needs and it tends to look like this.
| The message | What answering it needs | Send without a person? |
|---|---|---|
| Where do we park, and what time's check in? | The listing and the house guide | Often, once you've checked its answers for a few weeks |
| Could the sofa bed be made up? | A reply, and a job for the cleaning team | Only if the job lands as well as the reply |
| Can we stay two more nights? | The live calendar, and whatever your team agreed today | No. It's a booking change, so it's a team call |
| An Airbnb guest asks to extend by booking direct | Airbnb's off platform rules, which cover extensions | No. The helpful link breaks the platform's policy |
| The heating isn't working | How serious it is, who's on, and a line to hold the guest | No. It needs an owner before it needs a reply |
| Can we have a refund for last night? | Your authority line | Never. Anything that moves money stays with a person |
Only the top row is really a question. Everything under it is a decision with a message attached, and if the only thing you check is how well the reply reads, all six will pass.
Who owns the message comes before what it says
The fix isn't a better writer. It's deciding, before anything gets drafted, who should own each message: the AI, a VA, whoever's on ops, or you. That's the first thing OpSmithOs does with every message on our own guest inbox, across Hostaway and guests' WhatsApp.
Put the two extra nights back through it. Before it writes anything, it reads what your team agreed in Slack that day and what your maintenance contact said on WhatsApp, finds the shower booked in, and marks the message as the team's, because new dates are a team call anyway. Your VA opens it with the clash already spelled out and the next steps written, and there's a holding line ready so the guest isn't left waiting while someone decides.
Plenty of messages are simpler than that. A sofa bed request it keeps, drafting the reply and writing out the job for the cleaning team. A problem in the property comes back as an escalation that says how serious it is and who does what.
What shadow mode is, and what to mark it on
Shadow mode is what makes any of this safe to try. The AI reads every message as it lands and writes down what it would do, the reply it would send or the escalation it would raise. Your team carries on as they do now, and nothing reaches a guest on its say so. Each draft sits next to what your team actually sent, so you're judging it on your own guests in an ordinary week rather than on a demo.
You're not marking whether it sounds nice. Ask these of every draft instead.
- Would your best VA have sent it, word for word?
- Did it give the message to the right owner?
- Did it know what your team agreed earlier that day?
- Did it stay inside what it's allowed to offer?
- When it couldn't answer, did it keep the guest warm while a person decided?
Your team gets something from the first day, long before anything's trusted to send. The VA on shift opens a message that's already been read, given an owner and drafted, with the next step written out, so the job is checking and sending rather than starting from a blank box. And you get a record of every call it would've made, which is the only honest way to find out whether it makes them the way you would.
That's a real cost. It's also the reason nothing goes to a guest that your team wouldn't have said.
Whatever it sends, you've said
There's a reason to go slowly that has nothing to do with how good the AI is. In February 2024 a tribunal in British Columbia ordered Air Canada to compensate a passenger who'd relied on wrong advice about bereavement fares from the chatbot on its website. Air Canada had argued it couldn't be held liable for what its chatbot said. The tribunal saw that as treating the chatbot as a separate legal entity, responsible for its own actions, and rejected it.
It makes no difference whether the information comes from a static page or a chatbot.
It's a different industry in a different country, and none of this is legal advice. But your guest won't draw a line between your VA, your AI and you either. A reply that promises a late checkout, a refund or two free nights is your promise the moment it's sent, which is why the bottom of that table stays with a person even after everything above it has been trusted.
When should it send on its own?
It's ready when you'd sign off its drafts for one kind of message without reading them, because you've already read enough of them to know. You make that call from the record, one row of the table at a time, starting at the top. Parking and check in questions might get there in a few weeks. Booking changes might never need to, because a draft that's ready the moment someone picks the message up already saves most of the time.
It's how you'd bring on a new VA anyway. You'd read their replies before you let them send any, and you'd widen what they handle as they earned it. The AI doesn't get to skip that because it types faster.
Questions operators ask about this
Can AI reply to Airbnb guests automatically?
Yes. Some PMS and guest messaging tools can be set to send AI replies without anyone approving them, often for specific kinds of question rather than everything. Whatever goes out is sent from your account, though, so Airbnb's rules apply to it exactly as they do to your team, including the ban on links that take a guest off the platform. Whether it should send on its own is a separate question, and it's worth answering one kind of message at a time.
What is shadow mode in AI guest messaging?
It's when the AI reads every guest message and drafts what it would send or escalate, but nothing reaches the guest unless a person sends it. Your team works as normal, and each draft can be compared with what they actually did, so you learn where it's right and where it isn't before it's trusted with anything.
Which guest messages should AI never send without a person?
Anything that moves money, like refunds, discounts and deposits, and anything that can't be undone. Booking changes and complaints should go to a person too, because they need an owner and a decision before they need a reply.
Should I tell guests when AI is replying to them?
In shadow mode a person reads and sends every reply, so your guest is hearing from your team. Once anything sends on its own it's your call, and it's better decided before it goes live than after a guest asks.
Will AI guest messaging replace my VAs?
Not in shadow mode. It changes the job from writing replies to checking, deciding and sending them. The calls that need judgement still come to a person, they just arrive with the next steps written out.
Sources
- BBC Travel, Airline held liable for its chatbot giving passenger bad advice (23 February 2024)bbc.com
- The Air Canada chatbot decision, 2024 BCCRT 149, on CanLIIcanlii.org
- Airbnb Help Centre, Off-Platform and Fee Transparency Policyairbnb.co.uk
- Hostaway support, Inbox AI Replies: How It Workssupport.hostaway.com
Last checked . Vendor features change. If anything here no longer matches the vendor's own documentation, theirs is the one to trust.