Every tech company selling to hotels right now is selling AI. Robot concierges. Automated check-in kiosks. AI-generated guest messages that somehow feel less personal than a form letter.
Most of it is theatre. Here's what actually works.
The robot concierge problem
The robot concierge narrative has been around for fifteen years. The premise: guests want to interact with technology instead of people, so we should replace the human interaction with a digital one.
The data has consistently failed to support this. Guests who choose hospitality over Airbnb are specifically choosing a service environment. They want human interaction to be available and reliable, not absent.
Where AI excels is where it removes the friction that makes human service feel bad. Not replacing the service. Removing the friction.
Predictive maintenance: where the money is
The highest-ROI application of machine learning in hotel operations isn't guest-facing. It's maintenance forecasting.
A hotel has hundreds of mechanical assets (HVAC units, elevators, pool equipment, plumbing fixtures), each on its own failure curve. Reactive maintenance (fix it when it breaks) is expensive: emergency labor rates, guest disruption, potential brand-standard violations, and in some cases, PMS revenue loss while a room or amenity is taken out of service.
Predictive maintenance uses the operational history of each asset (cycle counts, service intervals, failure history, similar assets across the portfolio) to score each asset's failure probability before the failure occurs. A filter that's overdue by two cycles at a property where similar filters fail at four cycles is a maintenance ticket that should be filed today, not after the guest in 312 calls to complain about a musty smell.
This is not speculative; it's the same approach that power plants, aircraft fleets, and semiconductor fabs have used for decades. Hospitality came to it late, mostly because the operational data wasn't being captured in a structured way. That's changing.
Labor forecasting: the second-biggest lever
Labor is 35–40% of a hotel's total operating cost. Most properties staff by intuition and historical averages. The GM knows that Thursdays are heavy and Tuesdays are light, so they staff accordingly.
But the signal is noisier than intuition suggests. A Tuesday in late September when a conference comes to town is not a typical Tuesday. A Thursday the week of a major local event is not a typical Thursday. And micro-fluctuations (early checkouts, late arrivals, the group that changed their departure pattern) affect labor needs at a granularity that "average Tuesday staffing" can't capture.
ML-driven labor forecasting ingests the same signals a good GM uses (reservation patterns, departure curves, historical event data, weather) and produces staffing recommendations that update continuously as the day's actual pattern diverges from the forecast. You're not replacing the manager's judgment; you're giving the manager better data faster.
Properties that have implemented data-driven labor scheduling consistently report 4–8% reductions in labor cost per occupied room. At a 150-room property, that's not a rounding error.
Where AI fails in hospitality
It's worth being honest about where the current generation of AI tools underperforms, because the sales pitches aren't honest about it.
Guest messaging automation is oversold. LLM-generated responses to guest messages can handle FAQ-style queries adequately. They fail on anything that requires property-specific context, judgment, or empathy. The failure mode (a response that's technically coherent but obviously wrong) is worse for trust than a delayed human response.
AI check-in consistently shows lower satisfaction scores than human check-in at the same property. Guests tolerate it; they don't prefer it. The value proposition is cost reduction at scale, not guest experience improvement.
Revenue management AI is genuinely useful but only if the input data is clean. Garbage-in-garbage-out applies at full strength. A revenue management algorithm fed inaccurate OTA data or misclassified room types will confidently produce wrong recommendations.
The actual question to ask
Before any AI investment, the right question is: where are my humans spending time on work that machines are better at?
If your maintenance team is spending 30% of their time on reactive fixes that better scheduling would have prevented, that's where AI earns its money. If your GM is spending two hours every morning manually assembling the night audit from PMS exports and group chat summaries, that's a target.
The theatre version of AI solves a problem that doesn't exist (guests who want to interact with robots) and ignores the problems that do (information bottlenecks, reactive operations, labor inefficiency).
The productive version makes your humans faster at the things that matter.
The capabilities behind this dispatch
Where the ideas in this piece become day-to-day operations.