The wrong question
Most conversations about AI in hotels start with a framing that doesn't match how the technology actually gets used on property: will AI replace the front desk, the engineer, the duty manager. In practice, almost no hotel deploying AI today is trying to answer that question. The useful question is narrower and less dramatic: for this specific task, which part should a system do, and which part should a person do.
That distinction matters because guest-facing operations are made up of dozens of small tasks with very different shapes. Some are high-volume and pattern-based. Others are low-volume and depend on judgment, tone, or authority a system doesn't have. Treating all of them the same, either by automating everything or by resisting automation everywhere, wastes the strength of both sides.
Guest messaging: drafting versus deciding
A mid-size property can field 150 to 300 guest messages a day across SMS, WhatsApp, email, and in-app chat, covering everything from "what time is breakfast" to "I need to change my flight and my checkout." The first category is where AI earns its keep: answering repetitive, factual questions instantly, in the guest's language, at 2am when the night auditor is also covering the desk.
The second category is where a person needs to stay in the loop. A guest asking about checkout logistics tied to a flight change isn't really asking a factual question, they're signalling stress, and the reply needs judgment about fees, availability, and tone that a template can't carry. A workable split looks like this:
- AI drafts and sends responses to routine, low-ambiguity requests automatically.
- AI drafts a response for anything touching money, complaints, or special circumstances, and routes it to a staff member for a ten-second edit and send.
- Anything flagged as upset, urgent, or VIP goes straight to a person with the AI's summary attached, not a suggested reply.
The point isn't to minimize AI involvement in the hard cases. It's still useful there as a summarizer and context-gatherer. It just shouldn't be the one making the call.
Task triage: the dispatcher model
Inside the property, requests generated by messaging, PMS flags, or housekeeping apps need to land on the right person's list within minutes, not be rediscovered an hour later in a shared inbox. This is where AI functions less like an assistant and more like a dispatcher, scoring incoming tasks by urgency and routing them to whoever is actually free, rather than whoever happens to be on a static assignment sheet.
A platform like Hermes, which connects directly into a property's PMS including Oracle OPERA Cloud, is built around exactly this layer: it reads what's happening in the system, in guest messages, and in request logs, then routes the task and surfaces the context a staff member needs to act on it in the first ten seconds, not the first five minutes of reading history. The staff member still makes the decision and does the physical or interpersonal work. The system's job is just to make sure the right task reaches the right hands with enough context attached.
Illustrative example: a property running manual triage might see average time-to-acknowledge on maintenance requests sit around 25 to 40 minutes during busy shifts, simply because requests queue in a shared inbox. Automated routing by urgency and staff availability can plausibly cut that to single digits, not because the work got faster, but because the handoff did.
Service recovery: pattern detection, human delivery
Service recovery is the clearest case for collaboration rather than automation, because the two halves of the task genuinely require different capabilities. Spotting that a recovery situation exists, a late delivery, a noise complaint, a billing error, a guest who has now contacted the property three times about the same issue, is pattern recognition across disconnected systems. That's a task AI does reliably and faster than a person scanning five different logs.
Delivering the recovery is a different skill entirely. It requires authority to make a judgment call on a comp or rate adjustment, reading the guest's mood correctly, and in many cases simply showing up in person. No guest wants an automated apology for a bad night's sleep.
The useful framing one hotel operations lead offered during a product conversation: let the system tell me which fires exist before they spread, and let me decide how to put them out.
In this model, AI's contribution is almost entirely upstream: flagging the pattern early enough that recovery is still possible, and pulling the guest's stay history so the staff member isn't recovering blind.
Forecasting: better guesses, not fewer decisions
Demand and labor forecasting is often sold as a replacement for manager judgment, but the properties getting real value from it use forecasts as an input a human still interprets. An AI model trained on booking pace, local events, weather, and historical patterns can flag that Thursday is likely to run 20 percent above the usual midweek occupancy. What it can't do is decide whether to call in a part-time housekeeper, shift a staff member from a slower outlet, or accept slightly longer turnover times, because that decision involves budget constraints, union or contract rules, and knowledge of which staff are actually available that week.
The forecast shortens the decision, it doesn't make it. A manager who used to spend an hour reconciling three spreadsheets to guess at Thursday's staffing now spends five minutes checking a number against what they already know about their team.
The actual shift
None of this is about AI quietly taking over hotel jobs. It's a reallocation of which parts of each job are worth a person's attention. The repetitive, high-volume, pattern-based layer moves to software. The judgment, authority, and human-presence layer stays exactly where it was. Properties that get this split wrong in either direction, automating the apology or manually triaging the inbox, end up with worse guest experiences than properties that never tried AI at all. The ones getting it right simply have more hours in the day for the parts of the job that actually needed a person.