Riding an AI pilot takes a week, but taking him to the hotel operation takes much more

Riding an AI pilot takes a week, but taking him to the hotel operation takes much more
CECILIA RODRÍGUEZ SUÁREZ

CECILIA RODRÍGUEZ SUÁREZ

17 Nov 2026

Reading time: 5 minutes

Today any hotel can try an AI assistant or agent in a few days. What separates that test from a tool that the team uses every morning has to do with the data, the permissions, the costs and the way of working.

That's what our Chief AI Officer, Raúl Blanco Díaz, spoke about in the webinar "Where does an AI pilot end up and the transformation really begins?" Amadeus Company), and Lesmes López, Courance Regional Manager.

Watch the full webinar

Watch the full webinar


A prototype teaches little of what is below

When Lesmes López asked what distinguishes a company that tests AI from another that already obtains results, Raúl responded with data, simple and direct: today anyone can mount a small control panel with an AI tool in an afternoon, but that does not say anything. What decides if it reaches production is how the information with which you work has been collected, cleaned and processed. If the data is wrong, the AI amplifies the error; If they are cared for, it helps to decide better.

That is why Raúl insists on limiting each pilot to a specific area.A prototype that tries to cover everything is more difficult to evaluate and easier to abandon.The failures that you see the most in the projects that do not go ahead have less to do with the code and more to do with the security, the costs of using the model that no one calculated and the ignorance of a technology that changes all the time.

At NeoBookings we continuously measure what each model consumes in the products that are already in production, because what is implemented one day can be obsolete the next.


What happens when the information ismessy 

In a hotel, the information lives in the PMS, in the reservation engine, in the CRM and in the billing tool, and each system keeps it in its own way. If the language model receives that untreated data, it is more wrong, invents answers and spends more than necessary in each query. Raúl compares the data ordering work to giving him the food "already chewed, as if it were a porridge." 

To solve it there are two paths.OnneobookingsconsidersmosIt is necessary to order the data and build an agent-ready infrastructure, connecting the systems using MCP (Modelcontext PROTOCOL), the protocol that allows AI models to connect with other tools, bringing information to a«data warehouse»and«cleaning it»before the model uses it.“The complex path, which is what to take for me, is, first of all, ordering the data and making an architecture and infrastructure prepared for agents and LLMS”, ,explained. 

The second path is within reach of a hotel that cannot yet tackle that work by knowledge or by cost. A small agent can download PMS and CRM reports every day or week, process those four or five Excel, and return statistics and forecasts.It is a less complete solutionbutEven so, it improves the operations from the first day. 

Xavier VenteoHe added a nuance on which stages can be jumped. A hotel that already receives reservations through an API can build a conversational layer over it for the client to book from an assistant likeChatGPTorCopilot, and in that case the quality of the data weighs less. As soon as it wants to offer value-added services, recommendations or information that the guest is going to take as true, the data is once again essential.


Of the chatbot that answers the agent that it proposes

Raúl explained the difference between the chatbot of a lifetime, which loads frequently asked questions and answers, and an agent supported by a good language model, who consults information, uses tools and executes actions. Without deifying them, because in the end, he said, "they are still systems with instructions and hands to execute actions."

The example he gave was that of a hotel manager who arrives in the morning with fifty open fires. An agent can collect these incidents, order them by priority and propose how to attack them, with buttons for other agents to contact each department. If the fire extinguisher's technician arrives ten minutes late and solving it means sending him a WhatsApp, the agent takes care of it and returns a response so that the director decides. His time is left for what he really needs his expert and human judgment.

That is why Raúl believes that each company, small or large, will need someone who knows how to set up and shape these agents. Xavier, on the other hand, led him to governance. Without a group or a person who controls the process, AI multiplies organizational errors, it can compromise personal data of customers and shoots a budget that is no longer cheap, because the rates of the models are rising and it is paid per use.


The more power an agent has, the more you have to set your limits

Consulting available rooms has a very different risk from modifying a reservation or touching personal data. Xavier recalled that there have been cases in which an agent decided to delete an entire database, and hence the need to define what he cannot do and at what times a person has to give the go-ahead.

Raúl said that as part of the development of the products that we are carrying out, we dedicate a lot of iteration to limiting agents with guardrails, which are those rules that delimit what they can do.

Even so, he acknowledged that it is difficult to guarantee total security, because tomorrow there can be a question that no one had foreseen, and that the person continues to be the one who best detects the mischief of those who try to jump the limits. In this regard, Lesmes proposed to pass the Battery Assessment Assistant periodically and check that their answers remain within what was expected. Raúl's conclusion was to work in very closed environments and open little by little: "What is happening is that people deploy prototypes, tools, but with their doors open to the field, and that is where problems really arise"


The payment, the most delicate part

When an agent goes from recommending to collect, the data on the card and the PCI DSS come into play, the security rule that regulates any payment by card. Xavier explained that today these data should not go through a model housed outside the infrastructure itself, and less outside the European Union, where the regulation is more demanding. Where you see an immediate use is in helping you decide, for example, with which card less commission is paid or which option offers the best exchange rate in an international payment.

Raúl affirms that the agent payment will arrive and imagines it as a kind of natural language contract, of the type "you can buy these shoes when they fall from 50 euros", which the agent will execute when certain conditions are met. But for it to be a realityA standard protocol is missing that today does not yet exist.


When to climb and when to stop

Xavier is cautious and asks to make the bill before he grows up. With AI the capacity limit disappears, and with it the ceiling of the cost also disappears.It is necessary to project what happens if the service goes from one hundred thousand to a million customers and check that the income continues above spending.

Raúl agrees to measure, and defends that it is also leftMargin for small pilots whose profitability still cannot be calculated. Sometimes the only way to have data is to launch them, see how people respond and decide later if they deserve a professional version.

By accelerating, in addition, the bottlenecks change places. Xavier said that on his team the Product Owner validated everything before uploading it to production; When the AI multiplied what was delivered, it stopped coming, and they had to create a validation agent that only passes the cases that need its review. The same thing happens in a hotel.Automating the reception of a request is of little use if you then wait in three departments that are not coordinated.


Where to start

Raúl's advice for any hotel is to sit down with an AI tool and tell you what the operation is like for each day, step by step. From there you can draw patterns, detect what is repeated and ask for examples of agents that automate it. With a condition that you repeated several times, do it from a corporate account and never from a free one, so that the information of the hotel does not end up serving to train third-party models.

Choose this week a task that you repeat each day, describe it to the AI and measure how much time it returns to you. And if you need implementation and guide, we invite you to a coffee.

Contact Us