Table of Contents

In the world of technology, some innovations disappear (I haven’t sent a fax in a rather long time), while others persist (I do still ride my bike every day). The mistake, in the tech world right now, would be to believe that generative AI belongs in the persisting category capable of replacing all other forms of AI. A grave mistake.

In our article “Sophism and AI” (1) published in February 2024, we already predicted that all hotel projects based purely on generative AI would fail. Let’s check where we are with that today.

Let’s imagine you were appointed AI project manager for a hotel chain, and your task is to deploy AI for customer relations. Two options are available to you: classical conversational AI or generative AI.

Let us examine the strengths and weaknesses of each solution.

First of all, let’s recall the difference between intelligence and knowledge. Intelligence allows you to understand, analyse, and adapt, while knowledge encompasses the facts and information acquired through learning. In other words, intelligence understands the question, and knowledge provides the answer. In AI, both are essential. However, as intelligence becomes more of a commodity (Deepseek is just the beginning), knowledge becomes the real asset of these models.

Classical AI vs. Generative AI

With classical AI, everything must be designed from scratch: creating a comprehension model (NLU) and structuring the knowledge through a library or database. This requires anticipating all possible questions and answers, which is a long and costly process.

Conversely, generative AI simplifies everything. Intelligence is already integrated into models like OpenAI, Mistral, or Anthropic. For knowledge, all that is needed is to ingest existing data (web, technical sheets, PDFs, etc.) to build an exploitable base.

During a conference with hospitality professionals, I asked: Which AI would you choose for a hotel group? Result: 100% chose generative AI. Knockout victory? Unfortunately not, this was just the start of their issues…

You’re hallucinating, my friend

Generative AI is probabilistic and has a disturbing tendency to invent answers: this is known as hallucination. These errors have two main causes: a lack of data or, on the contrary, an excess of poorly managed information.

At a minimum, the hallucination rate is 20%. Are you ready to take this risk? Announcing non-existent accessible showers, offering fictitious late check-outs, or guaranteeing a heated pool… only by the sun?

Boolean Logic vs. Fuzzy Logic

Generative AI predicts each word based on the highest probability. But probability is not certainty. It excels at general queries (a pancake recipe, a marketing plan), where approximate answers can be adjusted by the user. This is fuzzy logic: adaptable, but not infallible.

Conversational AI, on the other hand, relies on boolean logic: true/false, black/white. Hospitality is full of boolean information:

Check-in is at 3 PM, not 2 PM.
The pool is 20 meters long, not 19.
The breakfast is gluten-free.
The showers are wheelchair accessible.

In short, how can an AI based on fuzzy logic manage an activity that requires boolean logic? It’s like trying to fit a circle into a square.

I then ask my audience of professionals once again. Suddenly, generative AI is now convincing only 50% of the participants. The initial enthusiasm has thankfully given way to some reflection.

Data, more data, always data

As mentioned earlier, a hotel—and even more so a hotel group—relies on a series of boolean information. We estimate that a hotel is characterised by around 3,000 data points:

What is the restaurant’s dress code?
Can the windows be opened?
What is the capacity of the safe?
Does the Wi-Fi cover the beach area?
Can the hotel provide a bottle warmer?

These examples are not chosen at random: these are very real examples of questions, and the response data simply does not exist in a formalised form. At best, the information is stored somewhere in the mind of that member of staff.

Our calculations show that 60% of a hotel’s information is undocumented. I repeat: it does not exist.

So, how can the best generative AI in the world answer a question it doesn’t know the answer to?

Now you understand and appreciate why knowledge is far more important than intelligence.

Maths is unforgiving

Why all the hype around generative AI? The answer lies in a lack of understanding of basic maths, specifically the logarithmic curve.

This curve teaches us that at first, progress is fast and inexpensive. But each marginal improvement becomes more complex and costly. In other words, the problems to be solved evolve exponentially.

What seemed simple at first eventually becomes an insurmountable puzzle. As illustrated by the curve below…

In the first year, you manage to get your AI to answer 40% of the questions. You think you’re a genius—another year, and you’ll have it all sorted.

Over the next two years, despite all your efforts, a larger team, and more spending, you barely surpass 60%.
By the fourth year, you’re asking for even larger budgets to gain just 5-10% more… In fact you’ve probably just been fired, and the project is definitively abandoned. That’s how millions of dollars and years of work are lost.

100% of hotel projects based solely on generative AI have failed

I ask the question for the third and final time. Now, finally, no one wants generative AI for their hotel project. In 15 minutes, they understood what large hotel groups, despite advice from top consultancies, didn’t manage to grasp in over two years.

And finally they’re right. 100% of hotel projects based solely on generative AI have failed, whether managed internally or entrusted to well-paid experts. Why? Because they focus only on intelligence, and they neglect knowledge.

The most surprising part? They do know. We’ve told them (and we’ve even written it down for them)

This logic also applies to AI Voice Bots: behind the attractive demonstrations lies a much less convincing reality, where the majority of responses completely lack sense and meaning. In any case, please remember that the best way to manage a phone call is to ensure that no one needs to make one in the first place.

Conclusion

According to NTT Data, between 70 and 85% of generative AI deployment efforts do not meet expectations. Forbes reported that 85% of AI models fail due to poor-quality data or a lack of data. And MIT reinforced this, saying only 5% of companies gain financial benefits from their AI technologies (see article(2)).

The problem is becoming so important that even the New York Times is talking about it in the article

“AI is getting more powerful, but its hallucinations are getting worse“.

So, what do you need to do? Bad news: you need to work. A lot. No pain, no gain.

But let us help you. Let us remind you that the menu we offer you here at Quicktext, hasn’t changed:

Starter: A solid team of (over 30) data analysts and data scientists to validate, modify, and evolve the model.
Main course: Knowledge: An established, full structured database, like Q-Data from Quicktext which includes 3100 data points which we are constantly adding to.
Dessert: Intelligence: A fully developed hybrid AI like Q-Brain+, 80% conversational, 20% generative.

Result: 90% of relevant answers with less than 1% hallucination.

Estimated time it would take you to create all of this from scratch ? At least 5 years. So maybe you should start now.

Or, if that sounds like an eternity and you want to move faster (3), come and see us…

(1) Sophism and Artificial Intelligence : https://www.quicktext.im/blog/sophistry-and-artificial-intelligence/

(2) 95% of generative AI pilots at companies are failing : MIT report: 95% of generative AI pilots at companies are failing | Fortune

(3) How Quicktext deployed AI for Medplaya’s 17 Hotels in 3 weeks: https://www.quicktext.im/blog/medplaya-quicktext-ai-partnership-in-17-hotels/

(4) The AI industry has a huge problem: the smarter Its AI gets, the more It’s hallucinating: https://futurism.com/ai-industry-problem-smarter-hallucinating.

Hotels winning without OTAs – here’s how

Learn the 10 proven methods to increase direct bookings and reduce third-party dependency.

Download the free checklist