Models Are Weather. Your Data Is Climate.
Data is the foundation of AI in hospitality
Every model we have today will be obsolete in eighteen months, replaced by something better that costs less. That's not a problem; it's a forecast. Models are weather. They roll through, they impress, they get replaced. Your data substrate is climate, the slow, structural condition that determines what can actually grow.
Which is why the most important questions in an AI strategy have nothing to do with which model anyone uses. There are three of them, they're about your data, and they have to be answered in order, because each one is worthless without the one before it.
Can the AI reach the data? Can it understand the data? Is it authorized to act on the data?
Access. Understanding. Authorization. That's the story, and it's shorter than most strategy decks.
Can the AI reach the data?
Here's the trap: at a single property, this is a solvable problem. A motivated GM, a capable integration partner, a few months of work, and you can wire an AI into one hotel's PMS, and the demo will be glorious. That success is exactly what misleads people. It convinces the organization that access is a project, when the pilot has quietly proven nothing about the portfolio.
The existence of a portfolio doesn't multiply the project; it changes the species of the problem. A brand runs hundreds or thousands of properties across multiple PMS platforms, multiple versions of each, decades of acquisition-driven sprawl, and a layer of per-property customization on top of all of it. The industry has spent thirty years building those interfaces, one expensive, bespoke connection at a time, and that's precisely the point: bespoke doesn't scale.
This is how AI initiatives live and die in pilot purgatory. The deployment that dazzled at the flagship knows everything about that one hotel and nothing about the guest standing at the front desk of the other 1,200. That's not Guest Intelligence at scale; that's a well-read stranger with one local friend.
For a single property, access is a project. For a brand, access is infrastructure: a unified, real-time stream of operational events across the entire portfolio, independent of which PMS happens to sit underneath any given roof. If you don't have that, every AI initiative you fund is starting from the parking lot, at every property, every time.
Can the AI understand the data?
This is the question that feels solved once the data is flowing. It isn't.
Getting data out of a system and knowing what it means are different achievements.
Hospitality data is a museum of local decisions: the rate code that means one thing at your Florida properties and something else in EMEA, the user-defined field that one region uses for ETA and another uses for parking, each PMS emits its own special format and structure for the same event.
Humans navigate this with tribal knowledge. There's a person at every property who knows what UDF23 means. She's retiring in March.
Large language models are spectacular at language but dangerously confident with ambiguity. Hand one a raw payload full of unlabeled local convention and it won't say "I don't know what this field means." It will guess, fluently. At brand scale, fluent guessing is a problem.
The fix is decidedly unglamorous: normalization, canonical schemas, semantics that hold across every property and every source system. It's the least visible work in this industry and the most valuable. Nobody brags that they spent four years defining how to normalize a rate change or an early departure across every system in the portfolio.
But that sentence is the difference between an AI that's reliable across your portfolio and one that left the pilot team nodding and everyone else semantically confused.
Is the AI authorized to act?
Here's the frontier, and it's worth saying that nobody has this quite nailed yet.
Hotel systems typically authorize people. A badge, a role, a login at a desk. There is no native concept of "this agent, acting on behalf of this brand, may adjust charges under fifty dollars but may not touch the rate code, and every action is attributable and reversible."
Until that concept exists (scoped, auditable, revocable authority for software actors), AI in hospitality stays in the suggestion business. Useful, but a fraction of the value.
The suggestion business has a ceiling. An AI that can only recommend is an AI whose output still has to be triaged, verified, and executed by the same overstretched staff it was meant to force-multiply.
The value provided is real but limited by what it's allowed to touch. The transformation everyone is budgeting for lives on the other side of delegated authority.
Scoped authority is built from exactly what the first two questions produce: events you can attribute, and semantics that hold everywhere. You can't audit what you can't reach, and you can't scope what you can't define.
Delegated authority is a data problem before it's an AI problem.
The order of operations
Access without understanding moves the mess closer. Understanding without access is a schema with nothing flowing through it. Authorization without both is a liability with admin rights.
Everything else is procurement.