Generative AI terms for hotel teams: prompt, token, context and more

Generative AI terms describe how a text model is fed, billed and judged. A prompt is the instruction you type, a token is the fragment of text it is counted in, and the context window is how much text it can hold at once. Each entry here gives a plain meaning, a hotel example and a check.

A software quote lands on your desk with a line for tokens and a line for seats, and nobody in the meeting can say which one grows when the front desk starts using the tool. The words are not difficult; they are simply never defined in the room. After reading this, you will be able to place every generative AI term into one of four families, ask the right question about each, and read an invoice or contract line with an idea of what it commits you to. For the underlying mechanism, see our article on how generative AI works with words.

How should a hotel team read these terms?#

Read every entry the same way: what the word means, what it looks like in a hotel, and what you must check before you rely on it or pay for it. The words fall into four families: the text (prompt, system instructions, output), the bill (token, usage limit, seat), the answer quality (grounding, retrieval, source, temperature) and the model (language model, multimodal, open and closed, version). Knowing the family tells you who in the hotel owns the question: the bill belongs to the finance office, the text to whoever writes the templates, the model to whoever signs the contract.

None of these words describes intelligence in the human sense. They describe how text is fed in, counted, held in view and turned into an answer. A definition you cannot act on is not worth memorizing, so each entry ends with a check: a question for a supplier, a line to look for on an invoice, or a habit for the team. Where a word appears in a contract, the useful question is not what it means but what happens when the limit is reached: the cap, the notice period, the version change. The words that describe what a system does on its own, such as agent, human confirmation and audit record, are defined in a separate glossary, and if a term here sits next to a system name you do not know, the hotel technology glossary covers PMS, API and channel manager in the same plain style.

What is a prompt, and what are system instructions?#

A prompt is everything you put in front of the model: the instruction and the material it should work from. System instructions are the standing rules set above your prompt by whoever configured the tool.

The prompt is not only the question. If you paste a rate sheet and write "summarize this for the sales meeting", the sheet and the sentence are both part of the prompt. System instructions sit above every prompt in the session and usually carry the operator's rules: which language to answer in, what the tool must refuse, which tone to keep. Staff often cannot see them, which is why the same question behaves differently in two tools.

The output is generated text, not a record. Treat it as a draft until a person has checked the numbers, the names and the dates against the source. A saved prompt or template is a prompt the team reuses, and it drifts: the same template gives different results after a model version change, or when a colleague edits it. Keep templates in one place, with one owner. For example, a template that turns a guest complaint into a reply in the hotel's tone, with the guest's name and the facts of the stay typed in by a person rather than produced by the model.

What to check: whether you can read the system instructions, whether saved prompts are shared across the team, and whether the tool keeps your prompts at all. That last one is a policy question, not a technical one.

What is a token, and why does the bill count them?#

A token is the fragment of text a model reads and writes in, and it is the unit that usage is measured and billed in. A short word may be one token; a long word several.

Tokens are counted twice: on the way in (your prompt and any pasted document) and on the way out (the answer). A long document pasted for a short summary is charged mostly on the input side. Turkish builds long words from stems and suffixes, and a Turkish word carrying a whole English phrase can arrive at the model as several tokens. A team that rewrites its prompts from English into Turkish can therefore see the same job measured differently.

A usage limit is a cap on tokens or requests over a period. What matters is what happens at the cap: the tool stops, slows down, or bills the excess. Ask before the front desk depends on it. A seat is a named user licence, the other way these tools are sold. Seats are predictable and tokens are not, and many suppliers mix the two on one invoice.

What is a context window, and what happens when it fills?#

The context window is the maximum amount of text the model can hold in view at one time, measured in tokens. It holds the system instructions, the whole conversation so far and anything you pasted.

Everything competes for the same space. A long thread about a group booking slowly pushes the original room list out of view, and the model then answers from what remains. When the window fills, the tool does one of three things: drops the oldest material, summarizes it, or refuses and asks you to shorten the input. None of these announces itself clearly.

The practical rule is one thread per task. Start a new conversation for a new guest, a new group or a new month, and paste only the pages that matter. Pasting a whole contract or a full season rate sheet is usually worse than pasting the two pages the question is about, because irrelevant passages compete with the relevant ones. What to check: whether the tool shows how much of the window is in use, and whether it tells you when it truncates.

What are grounding, retrieval, source and temperature?#

Grounding means the answer is written from material you supplied rather than from the model's general memory. Retrieval is the step that finds the right passages first. Temperature is the dial that decides how much variation the model allows.

Retrieval works like this: before answering, the tool searches a document set such as rate sheets, fact sheets or policy pages, and pulls the passages that match the question. The answer is then written from those passages. A source is the citation the tool shows next to a sentence, and it is only worth something if you open it and confirm the passage says what the sentence claims. A source line can point at a real document that does not support the claim.

Temperature decides variation: low values make the same prompt produce nearly the same answer, high values produce variety. Use low for anything with numbers, dates or policy, and higher only for draft marketing text you will edit anyway. Many business tools do not expose the setting at all. Grounding does not remove errors; the model can still misread a passage it was given. The difference is that the mistake is checkable against a named page. For example, asking "what is our cancellation rule for a group of 20 in July" against the hotel's own contract file, with the clause quoted back, rather than asking what cancellation rules usually are. For the failure mode, see our article on why assistants invent facts and how to catch them.

What do language model, multimodal, open and closed models, and version mean?#

A language model is the engine that produces text one token at a time by estimating what comes next. Multimodal means it also takes images or audio, open and closed describe where it runs, and version is which build you are talking to.

The mechanism of a language model is next-token estimation over a very large body of text. It produces fluent text; it does not look anything up unless a retrieval step gives it something to read. A multimodal tool accepts a photo or a recording as well as text. Hotel uses include reading a photographed damage report or a scanned supplier invoice, and the output still needs a person to check it.

Open models have weights that can be downloaded and run on machines you or your supplier control. Closed models are reached over the internet at the provider. The trade-off is control and maintenance against capability and effort: an open model you host is yours to update, patch and keep running. Where the model runs decides where the text goes, which is a data question, not a quality question, and it belongs in the contract rather than in the sales conversation.

Version matters because providers replace models, and the same prompt can answer differently after an update with no change on your side. Ask for notice of version changes and keep a small set of real hotel questions to re-run afterwards. What to check in a contract: which model family is used, how much notice of a change, where processing happens, and whether your prompts and outputs are used to improve the model.

Where are the agent, safety and record words defined?#

The words that describe what a system does on its own, such as agent, assistant, tool use, human confirmation and audit record, sit in a separate glossary, because they describe behavior and responsibility rather than text. The agent glossary covers the difference between a system that answers and one that acts.

The hotel AI policy article covers what staff may paste into a chat tool, which is the practical side of the prompt question. Read the two glossaries together: this one tells you what the tool is doing with text, the other tells you who is responsible when it does something. Keep both lists where the team can reach them, and add the hotel's own words as they appear in your tools and contracts.

What to do next#

Take the next AI invoice or tool contract you receive and mark each line with one of the four families: text, bill, answer quality, model. Then write five questions your team actually asks, such as a rate question, a cancellation question and a housekeeping question, and keep them as your test set. The next version change or new tool is then judged against something real rather than against a demo.