What "AI-Powered" Actually Means on a Hotel Tech Invoice

The word "AI" on a hotel technology invoice covers at least four different products, and the price does not tell you which one you bought. Knowing the difference is the single most useful piece of due diligence in the category, and it takes about ten minutes.

The four things the word covers

Rules and automation. If-this-then-that. Decision trees. Robotic process automation. If the guest arrives after 10pm, send the late-arrival message. Predictable, auditable, cheap, and not intelligent in any learning sense. Still the backbone of reliable automation in most properties, and correctly so for deterministic work.

Predictive and optimisation models. Classical machine learning - forecasting, regression, optimisation solvers. Demand next Thursday will be high; raise the rate. Needs clean structured data. Usually explainable. Has the clearest line to revenue of the four, because the outputs map directly to RevPAR, occupancy and labour cost.

Generative AI and large language models. The layer guests actually talk to. Understands a question phrased any way, answers in natural language, translates, summarises. Feels human when done well. Introduces failure modes the others do not have - hallucination, tone drift, saying something plausible and wrong.

Agentic systems. Combinations of the above that complete a task end to end rather than answering a question. Check the calendar, hold the slot, take the details, open a ticket for the team.

A vendor selling any one of these can print "AI-powered" on the invoice. The costs differ by an order of magnitude. So do the risks.

The question that sorts them

Ask: what happens when it does not know?

A rules engine hits a dead end and shows a menu. A predictive model returns a lower confidence score. A language model does one of two things - escalates to a person, or improvises an answer - and which one it does is the entire product.

The follow-up: is it answering from my data or from a general model? A language model grounded in the property's own knowledge base answers what it was told and hands over the rest. One answering from general training will produce a confident, fluent, wrong check-in time.

Two questions. Ten minutes. Most of the difference between vendors is in the answers.

Where each one earns its price

Rules earn it on deterministic workflows - routing, confirmations, reminders, status triggers. They do not earn it as a guest conversation layer, because guests do not phrase things the way the tree expects, and the fallback is a frustrated person typing 0 for an operator.

Predictive models earn it wherever there is clean historical data and a decision with a number attached. Pricing, staffing, purchasing. This is where the defensible ROI case lives, because the inputs and outputs are both measurable.

Language models earn it on the conversation itself - the forty questions a day, the guest writing in Georgian, the request that does not fit a template. The return is operational load and experience, which is real and harder to count than a rate change.

Agentic systems earn it when the task genuinely needs several steps and the steps are well-defined. They are the newest and the most oversold; ask for a live demonstration of the full task, not a description of it.

What it should cost

Roughly: rules are cheap and should be priced as software. Predictive models are priced as analytics and justified on a revenue number. Language-model products scale with conversation volume, because the vendor's own costs do - inference is charged per token and WhatsApp per conversation.

So a language-model quote that does not ask about your monthly conversation count is a quote that will change. And a rules-based chatbot priced like a language model is the most common overpayment in the category.

The risk column

Rules have no surprise risk; they do what they were told, including when what they were told is now wrong.

Predictive models carry data risk - garbage in, confident garbage out - and explainability risk if the vendor cannot say why a number moved.

Language models carry the risk that matters most for a guest-facing product: saying something wrong, warmly, at scale. The mitigation is grounding in your data plus honest escalation, and it is the thing to test hardest before signing.

What to do with this

Next time a proposal says "AI-powered", ask which of the four. Then ask the two sorting questions. Then price it against the right column - software, analytics, or per-conversation - rather than against whatever the word made you expect.

The vendors selling the real thing will welcome the questions. The rest will change the subject.