Hotels need task-level AI governance, not blanket approval

Ninety-eight percent of respondents to the Mews Hotelier Survey 2026 reported using AI in the previous six months, but 41% had no formal policy for it. The same survey found that 83% trusted AI to support decision-making.

Separately, h2c’s AI Opportunity Study 2026 found 91% of participating hotel chains using AI, while only 28% had a company-wide AI strategy led by senior leadership. The surveys cover different populations, so their percentages should not be combined. Their shared gap is stark: adoption has outrun decision rights, error tolerances and ownership.

AI rarely arrived through governance. A revenue recommendation improved, gained a button and stopped asking for approval, then joined a property management system, customer relationship platform, housekeeping app and three chat widgets. Ownership of specific decisions becomes unclear.

The default answer, “human in the loop,” can become theater. Staff approve 12 late checkouts the model already handles correctly, or sign invoice matches that have not been wrong in three months. The framework calls that “friction with a budget line”: the old process remains, plus a new approval screen.

The opposite error is more visible. A bot promises an unavailable room upgrade, and an automated complaint sounds like insurance copy, while an employee still keys invoices by hand. Visible automation does not mean the back office has become efficient.

What does this change for a general manager?

It replaces blanket approval with three choices: AI decides, AI recommends, or AI stops for a person. Each task gets tested against three conditions.

• Reversibility: If the output is wrong, how hard is it to undo? A wrongly flagged rate-parity issue takes a glance to correct; a wrongly canceled reservation costs a guest.
• Cost of the mistake: Measure it in money and trust. A minibar posting error is cheap in both, while a romantic package sent to a guest whose partner has died is cheap financially and catastrophic for trust. That exposure requires a human closer.
• Data readiness: Does the hotel have clean, unified data for this task? If the property management, customer relationship and point-of-sale systems disagree, the system cannot decide; personalization belongs in recommendation until the data becomes reliable. That is a judgment on the data, not on AI.

Autonomy does not belong automatically at the front of the guest journey. When trust is at stake, a human must stay closer. The safest autonomous tasks are routine and reversible; emotionally charged guest moments require a stop.

An ordinary operating day makes the hierarchy concrete. Each handoff carries a different level of financial, operational and emotional risk.

• 6 a.m. — Decide: Night-audit exceptions are checked, cash and card totals reconcile against the property management system, and mismatches get flagged. Nobody needs to approve arithmetic.
• 7 a.m. — Decide: The housekeeping sequence follows early arrivals and late departures, while an online travel agency rate-parity scan covers every channel. A supervisor can re-sort the sequence in 30 seconds.
• 9 a.m. — Recommend: A weekend revenue recommendation arrives with a rate move. The data is good, but a revenue manager should own the reason because the mistake carries real money.
• 11 a.m. — Recommend: A repeat guest arriving tomorrow has a profile suggesting a suite upsell. Hold it until the profile data is consistent; half the time, the “preference” was a one-off booking made by an assistant.
• 1 p.m. — Stop: A guest is unhappy because a room is not ready. The system can retrieve the timeline, prior stays and available options in a second; a person delivers the apology and sets the recovery value.
• 4 p.m. — Decide: A linen supplier’s invoice matches the purchase order and delivery note. Post it.
• 6 p.m. — Stop: A cancellation cites a family emergency. A human decides the waiver, and the guest hears it from a human.

Fifty-nine percent of hoteliers said the front-desk welcome and check-in should remain human-led. That view was strongest among properties already using AI extensively: greater familiarity did not persuade them to surrender the guest-facing moment.

How should autonomy be earned?

Hart’s rule is blunt: “autonomy should be earned, not configured.” Start in recommendation mode and log every override. When the override rate falls below the general manager’s comfort threshold and stays there for weeks, promote the task to decision mode; if it rises, demote it.

Review the task list monthly, like a profit-and-loss statement. That creates a living operating document rather than an unread policy, and treats human review as a starting position, not a permanent address. Hart describes this as a rule being set for a first pilot, not a result already claimed.

AI should replace approval work, manual reconciliation and invoice matching where evidence supports autonomy. The property pays for the technology and again in staff time if every output needs a signature. Hart’s warning: the hotel would have “a very expensive intern” and pay twice for every task it touches.

For hotels moving from scattered pilots to governed workflows, bot count will not be the next test. Override logs must show that systems have earned autonomy without shifting trust-sensitive decisions onto guests. The goal is fewer unnecessary signatures, clearer ownership and a human close at consequential moments.