Hotels need an operating model before they need AI agents

Half the industry has adopted AI, yet fewer than one in ten hotels report real business impact. That is not a model problem. It is evidence that hotels are putting agents on top of operating environments that have never been connected, described, or measured well enough to support reliable action.

The State of Distribution 2026 covers 343 PMS vendors, more than 270 hotel brands, and over 58,000 properties. Its headline is hard to ignore: roughly 50% of hotels use AI, while under 10% see real impact. The gap comes back to fragmented data, undefined success measures, and closed integrations.

Hotels do not need to wait for a better AI vendor to fix this. They need an operating model that tells a system what the work is, where the data lives, who owns the decision, and how a property will know whether the result improved the P&L.

Adoption without impact is a foundation failure

An AI agent can summarize a report, draft a guest reply, or suggest a rate. It cannot reliably improve an operation when the guest record is split across systems, the relevant workflow lives in someone’s head, and nobody agreed on the measure of success before the contract was signed.

The distribution study identifies a particularly practical constraint: only 7% of PMS companies publish fully public API documentation. The remaining 93% have closed or partly closed APIs. That does not make AI impossible. It does mean the promise of an agent acting across reservations, guest preferences, service requests, folios, labor, and F&B is often much larger than the data path beneath it.

This is also why vendor ROI claims deserve a harder operating review. If a provider says its tool will reduce response time, improve conversion, or save labor, the hotel should be able to state the current baseline, the exact workflow affected, the owner accountable for adoption, and the financial measure that will settle the question. Without those four things, a pilot is a demonstration, not a business case.

The same study carries a warning for commercial teams. Booking and search platforms are reading guest reviews directly to answer traveler questions. A generic five-star review gives those systems little to work with. Specific detail about rooms, service, dining, or location is now part of how a hotel is represented in machine-mediated search. That is a real use case for AI-assisted review analysis, but only if the hotel treats review content as operational input rather than a score to defend.

Room 407 cannot be optimized until the work is described

The clearest explanation of the missing layer comes from Room 407 and hotel ontology. A room is not just a room number in a PMS. In Martin Soler’s example, Room 407 is a Deluxe King, takes 31 minutes to clean, connects to Room 408, has recently serviced AC, and sits 75 meters from linen storage.

Each fact is useful. Together, they are operational context.

A housekeeper also knows things that are rarely captured in a system: this room takes longer because of its layout, a piece of floor equipment is worn, or the standard sequence forces an unnecessary trip across floors. The executive housekeeper knows it because she sees the work every day. An agent does not know it unless the hotel has documented it, connected it to rooms, tasks, equipment, spaces, and standards, and kept it current.

An agent cannot improve a workflow that the hotel has never defined well enough for a new supervisor to run.

That is what ontology means in plain operating terms. It is a usable map of how the property works, not a technical exercise for its own sake.

Soler points to a concrete finding: housekeepers can spend 18 minutes per shift retrieving supplies. That is not an AI strategy. It is a friction problem. Once it is visible, an operator can change supply placement, revise routes, adjust task sequencing, or test staffing patterns. AI can help identify the pattern at scale. It cannot replace the work of defining what counts as wasted movement and what a good standard looks like.

This sequencing matters because autonomous action is the last layer, not the first. Data comes first. Reporting follows. Then insights. Only then is a property in a position to let a system take action with confidence.

A shared guest record is not background IT work

The First Central Hotel Suites Dubai rollout is more useful than most AI announcements precisely because it is unglamorous. The 524-room property replaced disconnected front-office and F&B systems with Shiji Daylight PMS and Infrasys POS across two dining outlets.

The result is one shared guest profile across check-in, room service, and restaurant transactions. Shiji Middle East VP Andreas Duerkoop said staff could respond faster with fewer manual steps, while First Central Cluster Marketing Manager Dr. Zakaria Abdelhai said the property had greater confidence in its data. Those are modest claims, and they are the right ones.

No AI feature was the point of the announcement. The point was a coherent record that departments can use. Personalization, forecasting, and service automation all depend on it. If the front desk sees one version of a guest, room service sees another, and F&B sees none, an agent will only make the fragmentation move faster.

This is where many groups misclassify the work. PMS and POS consolidation gets treated as an IT project with an implementation timeline. It is actually an operating project. It changes who enters data, which field is authoritative, when a profile is updated, what a manager can see during service recovery, and how much re-entry a team tolerates during a busy shift.

The right test is simple: can a manager trace a guest interaction across departments without calling three people or opening three systems? If not, the connected record is the first deployment.

Start with a workflow, an owner, and a decision right

Are Morch’s AI readiness framework has five parts: leadership direction, data foundations, team skills, workflow integration, and governance. That order is more disciplined than the usual request to evaluate several AI tools and choose a winner.

It also matches Mark Charlinski’s criticism that the industry confuses having AI with having a strategy. As Charlinski puts it, a system that predicts a rate is fundamentally different from a system that changes it without asking. One informs a revenue manager. The other requires a decision right, a guardrail, an owner, and a plan for the exception.

Employee resistance belongs in this assessment. Morch argues that resistance is often a signal of failed past rollouts or concern about job security, not simple aversion to change. Operators should take that seriously. A team that has spent years correcting bad integrations has earned the right to distrust the next tool that promises to remove work. The answer is not a more energetic launch meeting. It is a workflow people can inspect, clear escalation rules, and proof that the new process removes a real irritation.

Pick one workflow. Name the owner. Define the decision the system will support or make. State what it may never do without human approval. Then build the data and integration path required to run that workflow.

Measure friction before promising transformation

The business case for this work is not abstract. In Shiji’s analysis of technology friction, Alejandra Pueblita calculates that a 200-room hotel at 70% occupancy and $162 ADR produces $8.29 million in annual room revenue. A 1% efficiency improvement is worth about $82,700 annually.

That is the number to bring into an integration discussion. Not because every automation produces a clean 1% gain, but because it forces the right question: where is the property losing time, accuracy, and attention today?

Shiji’s examples are operational, not futuristic. American Liberty Hospitality improved service speeds by 10% to 20% and increased F&B revenue by 10% to 15% across five properties after mobile POS deployment. Grand Hyatt Singapore reduced in-room dining order errors from 5% to 10% to near zero and cut menu-update time from weeks to hours after digitizing operations across 699 rooms. These are outcomes operators can recognize: fewer errors, less re-entry, faster service, and staff attention returned to the guest.

Start there. Baseline the manual handoffs in one workflow this week: a room-service order, an out-of-order room, a late checkout, or a group-rooming-list change. Count the systems touched, the re-entries required, the exceptions, the minutes lost, and the person who cleans up the errors.

Then decide whether an integration, a process change, or an AI capability is the right intervention. The order matters. Hotels that build the operating foundation first will have agents that can act on something real. Everyone else will have another tool looking for work it cannot reliably understand.