Chatbot or AI agent: the difference that decides your sales
"We have a chatbot" can mean radically different things: from a button menu reciting opening hours to a system that closes bookings, charges and notifies the front desk. The difference is not cosmetic — it decides how many enquiries end in a sale and how many leak to an intermediary. Let's sort it out.
The difference, in five axes
The clearest framework was published by Hospitality Net (2025): an agent differs from classic conversational AI along five axes. Scope: single task versus multi-step processes. Autonomy: responds when spoken to versus acts to fulfil a goal. Memory: stateless versus persistent. Planning: none versus decomposing a goal into steps. Adaptability: fixed rules versus learning.
The pocket version: the chatbot tells you how to book; the agent books.
The three-level ladder
To self-diagnose what you have (or what you are being sold), the three-level ladder proposed by Neobookings (2025 — a direct-booking vendor, duly noted) works well:
- Informational assistant. Answers questions: hours, services, policies. Real but limited value: every transaction remains human.
- Scoped executing agent. Books, modifies, charges. The conversation becomes a purchase process.
- Operational orchestration. After the booking, it alerts housekeeping, notifies reception, updates the CRM. The system stops being a channel and becomes part of operations.
One caveat before the ladder: the rigid decision-tree chatbot — the button menus of a few years ago — doesn't even score. It locks up at any variation and is no longer considered a valid investment. The real entry point is conversational AI that understands intent, synonyms and typos.
The Spanish cases with numbers
This is not conference theory. RIU has operated its Claud·IA assistant since 2024 (built on Google Cloud and Gemini, with Emergya): it handles 75% of more than 1,500 daily queries, with a 73% booking completion rate among registered users, according to figures published by the chain itself. It is Spain's best documented case — level 2 of the ladder, achieved.
Eurostars (Hotusa Group, more than 300 hotels in 24 countries) launched a conversational booking assistant on its official website in 2026: users search and book in natural language without leaving the conversation. And Mirai operates Sarai, a system of coordinated specialised agents serving 100 languages by text and 33 by voice. From that last case, the most useful fact for any business is not technological: when connecting the agent, automatic loading from the website only covers around 45% of customer questions. The rest is internal knowledge that was never published anywhere. And almost 60% of the questions travellers ask are about the hotel, not about the booking: about what the guest will find on arrival, not about the act of booking (Mirai, March 2026 — a vendor, but the pattern matches what anyone who has staffed a bookings phone line will recognise).
The uncomfortable lesson: the bot is worth what your data is worth
From all of the above follows the conclusion almost no chatbot seller highlights: the system's quality depends less on the engine than on the documentation feeding it. If your cancellation policies live in the receptionist's head and your supplements in a 2019 spreadsheet, the best agent on the market will answer vaguely. That is why the industry's own recommendation is to start by ordering the knowledge base — what each rate includes, exact policies, edge cases — before choosing a tool.
My experience with my own prompt-improver tool (far more modest than a hotel engine, but with the same anatomy: a model + a knowledge base + usage caps) points the same way: 80% of the result is decided in the documents, not the model.
Where to start, depending on your situation
- You have nothing: start at level 1 with your 10 basic information blocks well written (rates, booking, policies, services, directions, deal-breaker questions). It is cheap and the payoff in staff time is immediate.
- You have a button bot: replace it; it is not a foundation for anything. Today's conversational standard understands natural language in dozens of languages.
- You have level 1 and volume: the jump to executor (level 2) is where the money is — every question resolved at 11 p.m. without closing the sale is a booking that probably ends up elsewhere.
- You are a chain or run several systems: level 3 demands integrations (PMS, CRM, engine) and governance: human oversight, traceability and clear limits on what the agent may do alone.
The right question when evaluating any offer is no longer "does it have AI?". It is: which level of the ladder is it on, what data feeds it, and what happens when it gets things wrong?
