The Human-in-the-Loop Hotel: Where AI Should Decide, Recommend and Stop

AI is already part of everyday hotel operations. In the Mews Hotelier Survey 2026, 98% of respondents reported using AI in the previous six months, 83% trusted it to support decision-making, and 41% had no formal AI policy. Separately, h2c’s AI Opportunity Study 2026 found that 91% of participating hotel chains were using AI, but only 28% had a company-wide AI strategy led by senior leadership.

These are different surveys, but they point to the same tension: adoption is widespread, while formal direction and oversight remain uneven. Hotels are putting AI to work. Whether they have clearly defined where it should decide, where it should recommend, and where it must stop is another question.

I do not say this to scold anyone. It is the natural result of how the technology arrived. Nobody signed a governance document before turning on a rate recommendation. The revenue tool got better, then got a button, then the button stopped asking. Multiply that by a PMS, a CRM, a housekeeping app and three chat widgets, and you have a hotel where no one can clearly say who is in charge of a particular decision.

The industry’s reply to this discomfort has been a phrase: human in the loop. I like the phrase. But it has quietly become an excuse, because in practice it often means human in every loop. A person approves the same twelve late check-outs the system got right yesterday and the day before. A manager signs off on invoice matches that have not been wrong in three months. The old process is still there, a machine has been bolted onto it, and someone has been given a new screen to click. That is not governance. It is friction with a budget line and it is how many properties will explain, two years from now, why nothing changed.

The opposite error is just as common, and more visible. Hotels race to put autonomy exactly where the guest can see it: the bot that promises a room upgrade it cannot deliver; the automated reply to a complaint that reads like it was written by an insurance company. Meanwhile, in the same building, someone is still keying invoices by hand.

So the real question for a general manager is not whether to keep humans involved. Of course you will. It is how to decide, task by task, whether AI should decide, recommend or stop and wait for a person. And the honest answer has very little to do with how clever the AI is.

It comes down to three questions about the task itself.

The first is reversibility. If the system gets this wrong, how hard is it to undo? A wrongly flagged rate-parity issue takes a glance to correct. A wrongly cancelled reservation costs a guest. Reversible tasks can tolerate autonomy long before their irreversible cousins.

The second is the cost of a mistake, measured twice: once in money and once in trust. These do not move together. An automated minibar posting error is cheap in both currencies. An automated message to a guest whose partner has just died, offering them a romantic package, is cheap in money and catastrophic in trust. Anything where the second currency is at stake needs a human closer to the decision.

The third is data readiness, and this is where much of the industry gets ahead of itself. Does the hotel actually have clean, unified data for this task, or does the answer sit across three systems that disagree? You cannot let a system decide when your PMS, CRM and POS are telling different stories. Much of what is sold as personalization sits precisely there. It is not that AI cannot do it; the data beneath it is not reliable enough to trust with a decision. Those tasks belong in the recommend bucket – not as a judgement on the technology, but on the data.

Run an ordinary operational day through those three questions and the buckets sort themselves. The result inverts the intuition that autonomy belongs at the front.

At six in the morning, night-audit exceptions have been checked, cash and card totals reconciled against the PMS, and mismatches flagged. Decide. It is reversible, cheap to get wrong, and the data is already structured. Nobody needs to approve arithmetic.

By seven, the housekeeping sequence has been drafted around early arrivals and late departures, and the OTA rate-parity scan has run across every channel. Decide, again. If the sequence is wrong, a supervisor re-sorts it in thirty seconds.

At nine, the revenue recommendation for the weekend comes in with a rate move attached. Recommend. The data is good, but the cost of a mistake is real money, and a revenue manager should still own the reason for the move.

At eleven, a repeat guest arriving tomorrow has a profile suggesting a suite upsell. Recommend and keep it there until the profile data is consistent. Half the time, the “preference” was a one-off booking made by an assistant.

At one, a guest at the desk is unhappy about a room that was not ready. Stop. The system can pull the timeline, previous stays and available options in a second. A person delivers the apology and decides what it is worth. 59% of hoteliers said that the front-desk welcome and check-in should remain human-led. The finding was most pronounced among properties already using AI extensively, which tells you something: the people who know the technology best are drawing the line hardest at the guest’s face.

At four, an invoice from a linen supplier matches the purchase order and delivery note. Decide, and post it. At six, a cancellation comes in referencing a family emergency. Stop. The waiver is a human decision, and the guest should hear a human make it.

Notice the pattern. The safest places for AI to act on its own are the boring ones. The places it should stop are the moments with a face on them. Most hotels are currently spending their courage in the opposite order.

There is one more rule that makes the whole thing workable on a Monday rather than a slide: autonomy should be earned, not configured.

Start everything in recommend. Log every time a person overrides the suggestion. When the override rate on a task drops below a level you are comfortable with and stays there for a few weeks, promote it to decide. If it climbs back up, demote it. Review the list monthly, the way you review your P&L. This gives a general manager a living document instead of a policy nobody reads. A human in the loop is a starting position, not a permanent address.

We are still choosing our first pilot to run this way, so consider this a rule we are setting rather than a result we are claiming.

Which leaves one uncomfortable question for anyone who has kept a person on every decision in the name of safety. If your AI needs approval for everything, you did not buy a colleague. You hired a very expensive intern, and you are paying twice for every task it touches.

Where, in your building, is the intern still waiting for a signature it stopped needing months ago?