AI agent, assistant or chatbot? A glossary for hoteliers

A chatbot answers. An AI assistant answers from your hotel's own data and can carry out routine actions, with a person confirming sensitive ones. An AI agent is software that acts toward a goal through tools and permissions. The words describe how much a system may do, not how clever it is.

A vendor opens a laptop, types a question in plain English, and a room rate appears with a confident explanation underneath. Before you decide whether that is useful, you need to know which of four or five different things you are looking at, and what each one is allowed to touch. This glossary defines about twenty words you will hear in one sales meeting, from chatbot to model-agnostic, and gives you the question to ask when each one appears. If you want the broader technology vocabulary first, see our hotel technology glossary: PMS, API, channel manager and more.

What is the difference between a chatbot, an AI assistant and an AI agent?#

A chatbot answers questions. An AI assistant answers from your hotel's own records and can carry out routine actions. An AI agent acts toward a goal through tools and permissions.

A chatbot replies from a script or from a general model with no access to your property. It cannot see whether room 402 is free tonight, so it can only talk. An AI assistant reads the hotel's own data, answers in plain language and performs a defined list of actions, with a person confirming anything sensitive. An AI agent is given a goal, a set of tools (check availability, place a hold, send a confirmation) and permissions. With each step up, the risk shifts from wrong words to wrong actions.

The vocabulary is unsettled even for regulators. Türkiye's Personal Data Protection Authority, in its document "Etken Yapay Zekâ (Agentic AI), Şubat 2026" (published 12 March 2026), notes there is no single agreed definition, describes systems made of AI agents acting with different degrees of independence, and uses the terms "Etken YZ" and "YZ Aracısı", adding that "YZ Ajanı" is also used. In Turkish, "yapay zekâ ajanı" means software that carries out actions on your behalf; in an English-language meeting, the practical question is the same: which of the three is it, and what is it allowed to touch? A vendor who cannot place the product in one row of the table below is describing a feeling, not a system.

The AI agent vs chatbot distinction is not about intelligence. It is about reach.

What do access words mean: action, scope, consent, revoking access?#

They describe what the system may reach and do, and who agreed to it. Access is the part of an AI product a hotel can actually control.

An action is one defined operation the system can perform, such as check availability, place a hold or post a charge to a guest bill. Ask for the list, not a description. Scope is the boundary of that permission: which rate codes, which date range, which outlets, read only or also write. Reading availability and changing a rate are different permissions; a system that can do the second should say so in the same sentence as the first.

Consent means a named person said yes to a specific access before it happened, and understood what it covered. A clause in a contract signed a year ago is not consent for today's connection. Revoking access should be a switch that works today, not a support ticket. Ask who can flip it, how long it takes and whether the change is recorded. For what your own staff may share with AI tools in the meantime, see our guide on hotel AI policy: what staff may share with AI chat tools.

What is MCP, in plain words?#

MCP, the Model Context Protocol, is an open agreement that gives an AI application a standard way to connect to outside data and tools, instead of a custom connection for every pair of systems.

It has three parts: the host (the AI application), the client (the connector inside it) and the server (the service that offers data and tools), per the Model Context Protocol project specification (opens in a new tab), protocol version 2026-07-28. A server can offer resources (context and data), prompts (templated messages and workflows) and tools (functions the model can run), the same specification says.

Why should a hotelier care? One well-built connection can serve several assistants rather than one connection per assistant. But the connection is only as safe as the permissions behind it. The specification's security principles say users must explicitly consent to and understand all data access and operations, and the application must obtain explicit consent before invoking any tool. Authorization is optional in the specification; where a web connection uses it, it builds on OAuth 2.1, the common standard for giving an application limited access on a person's behalf, and clients should request only the permissions the task needs, per the specification's authorization section (opens in a new tab).

The question to ask: "Which tools does this connection expose, and who approves each one?"

What do answer words mean: source, hallucination, handover?#

They describe where an answer comes from and what the system does when it cannot support one. An answer without a source is a guess with good grammar.

The source is the record the answer is drawn from: the rate table, the reservation, the hotel's written policy. Ask to see the answer with its source attached, not just the answer. Hallucination, which NIST's Generative AI Profile (opens in a new tab) (NIST AI 600-1, July 2024) calls confabulation, is "the production of confidently stated but erroneous or false content" that may mislead the person reading it. The same NIST document warns about automation bias, the tendency of people to over-trust a machine's output. A fluent answer is easier to over-trust than a clumsy one.

No error rate you are quoted can substitute for your own test. Ask your five hardest front desk questions and watch what happens when the system does not know. Handover is the defined moment the system stops and passes the conversation to a person, plus what that person receives: the transcript, the guest's name, and anything already promised. The question to ask: "Show me an answer your system refused to give, and what it did next."

For the wider picture of what these systems can and cannot safely do in a hotel, start with AI in hotels: a practical guide to uses, limits and safeguards.

What do safety words mean: confirmation, safety tier, two-person approval?#

They describe how much human agreement an action needs before it happens. The point is to spend human attention where a mistake is expensive.

A confirmation is one named person approving one specific action, having seen what the action will do: the room, the amount, the guest. A safety tier is a fixed grouping of actions by consequence: some run directly, some need one confirmation, the most sensitive need two. The tiers should be visible in the product, not buried in the contract. Two-person approval belongs on irreversible or high-value steps, for example a refund, a rate override or the deletion of a reservation.

The OWASP Top 10 for LLM Applications (opens in a new tab) (2025 edition) names "Excessive Agency" (LLM06) as the weakness that lets damaging actions follow unexpected, ambiguous or manipulated model output. It traces the weakness to too much functionality, too many permissions or too much independence, and recommends limiting tools and permissions to the minimum and requiring a human to approve high-impact actions.

Approval fatigue is the quiet failure: if everything needs a click, people click. Tiering exists so that a confirmation still means something. The question to ask: "Which actions need two people, and can I see that list today?"

What do record words mean: log, decision record, audit trail, tamper-evident?#

They describe what is written down after something happens, and how far you can trust the writing. A hotel needs records a manager can read, not only records an engineer can read.

A log is a technical list of events, useful for diagnosis and rarely readable by a general manager. A decision record is one plain-language entry for one change: what changed, why, who or what proposed it, who approved it, and whether it can be undone. An audit trail is defined by the NIST Computer Security Resource Center glossary (opens in a new tab), citing NIST SP 800-53 Rev. 5, as a chronological record of system activities sufficient to reconstruct and examine the sequence of events leading to an operation, from its start to its results. Tamper-evident is defined by the same NIST glossary, citing NISTIR 8202, as "a process which makes alterations to the data easily detectable". It does not mean the data cannot be changed; it means a change shows.

Check three things: can a manager read an entry without help, can it be exported for an accountant or an inspector, and how far back does it go. The question to ask: "Show me last week's entries for one rate change, and show me what happens if someone edits one."

What do evaluation words mean: test set, trap case, benchmark, model-agnostic?#

They describe how someone decided that a particular AI model is good enough for a particular job. Without that decision being written down, a demo is the only evidence you have.

A test set is a fixed list of questions or tasks with known correct answers, run the same way each time so that two models, or two versions of one model, can be compared. A trap case is a question built to catch a plausible wrong answer: a rate for a date with no availability, a service the hotel does not offer, a policy that does not exist. A benchmark is the test set plus the scoring rule. A score without the rule and the cases means nothing, and a benchmark that has not been run is a plan, not evidence.

Model-agnostic means the system is not tied to one AI model, so the model can be chosen per installation by measurement. Ask what happens when a better model appears, and what happens when a model is withdrawn. Start narrow: a back-office task with a checkable answer is easier to test than a guest conversation, and it teaches you how the vendor behaves when the answer is wrong. Our guide on where to start with AI in a hotel: back office before guests walks through that choice.

What to do next#

Take five questions into your next vendor meeting: which of the three is it, what may it touch, where does the answer come from, who approves the action, and where is it written down. The vocabulary fits together as one chain, from answering to acting to being accountable. A product that cannot answer the last three questions has only shown you the first.