There is a pattern in how new technology enters hospitality: a vendor ships a feature, hotels buy it because of the marketing, and then the feature sits unused because it doesn’t fit how the operation actually runs. I’ve watched this happen with mobile keys, chatbots, and now it’s happening with AI.
I built a dual-interface AI system a guest-facing concierge and a staff-facing operational assistant specifically to map where the technology provides genuine value and where it creates new problems.
Where AI genuinely helps
Guest FAQs and service requests at scale. A hotel with 200 rooms receives roughly the same 40 questions every day: check-in time, parking availability, breakfast hours, local restaurant recommendations, how to connect to wifi. An AI concierge can handle all of these without staff involvement, at any hour, in multiple languages. This is real value not because it’s cheaper than hiring people, but because it’s available at 2am when no one is at the desk.
Pattern detection in operational data. “Which room types had the most maintenance requests in the last 30 days?” “Are there any guests arriving today with a history of complaints we should brief the team on?” These are questions a good GM asks constantly. The data to answer them exists in the PMS and maintenance logs, but pulling it manually takes time. A well-designed AI assistant turns this into a conversation.
First-pass summaries of guest feedback. Reading and categorising 200 post-stay surveys is a task that takes hours. An AI can produce a first-pass summary in seconds. The summary still needs human review and this is important but it accelerates the process significantly.
Where AI breaks in hospitality specifically
Anything touching real-time availability. A guest asks “can I get a late checkout?” The AI doesn’t know occupancy for tomorrow. It doesn’t know if the conference group arriving at 3pm needs every room turned over. Unless the AI is connected to live PMS data and trained to understand the operational implications, it will either refuse to answer (frustrating) or give an answer that creates a problem for the front desk (dangerous).
Confident answers about uncertain things. This is the failure mode I see most often in hospitality AI demos: the system gives a confident, fluent answer to a question it doesn’t actually have grounding for. “Yes, we have a pool” but the pool is closed for renovation. Hallucinated or stale information in a guest context damages trust in a way that’s hard to recover from.
Anything that requires operational authority. An AI cannot and should not commit to a room upgrade, waive a fee, or make a service recovery decision. These require human judgment and often human accountability. The right design is always: AI presents options and information, human makes the decision.
The design principle that changes everything
Guest-facing and staff-facing AI have opposite risk profiles. A guest is comfortable hearing “I’m not sure let me connect you with someone.” It feels like good service. A housekeeping supervisor cannot afford ambiguity when they’re scheduling a floor for turnover.
Guest-facing AI should fail gracefully and hand off. Staff-facing AI should show its working and require confirmation for anything consequential. These are not the same product with different labels. They require different architectures, different confidence thresholds, and different failure modes.
What I’d build if I had access to a real PMS
The most valuable thing I couldn’t build in a demo: real-time PMS context injected into every AI interaction. Knowing that a guest has stayed four times before, always requests a high floor, and left a noise complaint on their last visit and having that context available to the concierge before the guest even asks a question. That’s where AI in hospitality becomes genuinely interesting. The technology isn’t the constraint. Access to the data is.