AI guest messaging: what it can and cannot handle
Automation is not a percentage of your inbox. It is a specific set of conversations, defined by what you have written down and what you have authorized.
The question this answers: Which guest conversations should stay automated, and which need a person?
7 min read
The question is authority, not difficulty
Most discussions about AI in guest messaging get stuck on how clever the model is. That is the wrong axis. A guest asking for the Wi-Fi password and a guest asking for a refund are not different in linguistic difficulty — they are different in who is allowed to answer.
The useful test has two parts. First: is the answer already written down and approved? Second: does answering commit you to anything — money, access, a change to a booking, a promise about the property? A conversation that passes the first test and fails nothing in the second can be automated safely. Everything else is a routing problem.
If the answer is not written down anywhere, automation is not the blocker. Documentation is.
The three routes, with worked examples
The examples below are illustrative rather than drawn from customer accounts. They exist to show where the line falls in practice.
| Guest says | Route | Why |
|---|---|---|
| “What time is check-in and where do we park?” | AI answers | Written down, approved, commits nothing |
| “Is the pool open in November?” | AI answers | A property fact, if it is in the approved knowledge |
| “Can we get in an hour early?” | Specialist | Depends on the day's occupancy and your policy on exceptions |
| “The door code isn't working.” | Specialist | Troubleshooting first, then escalation if it is a real fault |
| “This room is not what we booked.” | Specialist | Judgment, tone and de-escalation before anything else |
| “We want a refund for last night.” | Client decision | Money. Prepared and routed, never decided by support |
| “There's water coming through the ceiling.” | Client, by emergency route | Needs someone physically present, immediately |
What makes the first column bigger
The share of conversations that can be answered without a person is not fixed. It is a direct function of how much of your property knowledge has been captured, structured and approved — and that is something you control.
- Write down the answers you already give
- Most operators can reconstruct eighty percent of their repeat answers from a month of sent messages. That archive is the raw material; it just needs organising and approving rather than inventing.
- Be specific about exceptions
- “Early check-in is sometimes possible” cannot be automated. “Early check-in from 1pm when the previous guest has checked out, otherwise bag storage” can be.
- Keep it current
- A knowledge base decays. New appliance, new parking arrangement, new cleaning partner — each one silently turns a correct answer into a wrong one. Ongoing upkeep matters more than the initial build.
- Name what must never be answered
- An explicit deny list — refunds, compensation, medical, legal, anything about other guests — is as important as the approved answers, and shorter to write.
The part most comparisons skip: what happens at 3am
AI can respond around the clock within approved knowledge. Human availability is a purchased thing, and it is where models genuinely differ.
Outside the hours you buy, the honest behaviour is for automation to keep answering what it properly can, gather details on what it cannot, and queue that for the next coverage window — while stating plainly that a person will follow up rather than implying one is already reading. Emergencies are the exception and should never sit in a support queue at all: they need a predefined route to someone who can act.
A provider that cannot tell you exactly what happens in the hours you did not buy has not thought about the hours you did.
Managed guest messaging
How Relaytask runs this split as a managed operation — approved knowledge, trained specialists on the exceptions, and your authorization where it belongs.
FAQ
Common questions
Anyone quoting you a number before seeing your inbox is guessing. The realistic answer depends on how repetitive your guest questions are, how much of your property knowledge is documented, and how many exceptions your policies contain. Start by categorising a month of real conversations against the three routes above — that gives you a number grounded in your own operation rather than someone's marketing page.
They should never be misled into thinking a person is present when nobody is. The behaviour worth insisting on is that automation does not impersonate availability it does not have, and that a handover to a person is visible rather than disguised.
In a managed operation that is an operational event, not your problem to chase: the conversation is picked up, the knowledge or rule that produced the wrong answer is corrected, and the correction persists. In a software-only arrangement, that correction loop is work you own.