Scripted Bot or LLM Assistant: How to Tell What You Are Buying

Two products are sold as "hotel chatbots", and the difference between them is the difference between the technology that failed in 2019 and the technology that works now. You can tell which one you are looking at in three questions.

The one that failed

The first generation was a decision tree. Predefined buttons, scripted paths, if-this-then-that. It could answer what time is check-out if the guest pressed the right option. It could not answer can we check in early if we land at 8 because nothing in the tree matched.

Guests learned to avoid it. Staff stopped trusting it. Workload went up, not down, because the desk spent its time correcting, explaining and apologising for what the bot had said or failed to say. In an industry built on experience, that friction was fatal, and many properties quietly switched it off.

That product is still sold. It has a nicer interface now and sometimes the word AI on the box. It is still a tree.

The one that works

The current generation runs on large language models - the same technology behind general-purpose assistants. It understands meaning, tone and intent rather than matching keywords. A guest can phrase a question any way, in any language, with typos and two questions in one message, and it understands the whole thing.

Grounded in the property's own knowledge base, it answers what it was told and escalates the rest. It reads a voice note in Polish and replies in Polish. It recognises frustration and hands over rather than pushing through.

That is not a better tree. It is a different kind of product, and the difference is why properties that gave up on chatbots are trying again.

Three questions that sort them

1. Phrase a question badly. Type something like hi we land at 8am is early checkin possible also is there parking thanks. A scripted bot matches one keyword and answers one thing, or offers a menu. A language model answers both parts in one reply.

2. Ask something the vendor could not have prepared. Something specific to your property that is not in the demo script. A tree dead-ends. A language model either answers from the knowledge base or says plainly that it will connect you to the team. Watch for the third possibility - a confident answer that is wrong. That is the failure mode of the current generation and the one to test hardest.

3. Send a voice note in a language nobody in the room speaks. A tree cannot process it at all. A language model transcribes, understands and replies in that language. Have a speaker read the reply.

Three questions, five minutes. The demo cannot hide the answers.

Where the old kind still belongs

Rules are not obsolete. They are the right tool for deterministic work: sending the pre-arrival message at 9am the day before, routing a housekeeping request to the housekeeping group, triggering the departure reminder. Predictable, auditable, cheap.

The mistake is using rules as the guest conversation layer, where the input is human language and the tree cannot anticipate it. And the opposite mistake is using a language model where a rule would do - paying inference costs to send a confirmation that never varies.

Good products use both, and can say which is doing what.

The risk that comes with the new kind

A language model's strength is fluency, and fluency is also the risk. Answering from general training rather than your data, it will produce a check-in time that sounds completely right and is invented.

Two things mitigate it, and both are testable. Grounding: it answers from what the property provided, not from the model's general knowledge. Honest escalation: when it does not know, it says so and hands over instead of filling the gap.

Ask the vendor about both. Then test question two above, because that is where the answer shows.

What to buy

If a vendor's product is a tree, know that you are buying the thing hotels abandoned, and pay software prices for it, not AI prices.

If it is a language model, ask what it is grounded in and what it does when it does not know. Then run the three questions with your own property's information loaded, because a product that passes them in the demo and fails them on your data has told you what onboarding will be like.