Machine learning, deep learning and generative AI: the AI family explained for hoteliers
This guide explains the AI family for hoteliers. It shows how to test tools and ask the right questions before buying.
Contents
Key takeaways
- Artificial intelligence is the umbrella term; machine learning, deep learning, natural language processing, computer vision and generative AI are different ways of building something under it.
- A product can be sold as AI while the actual mechanism is a fixed rule a person wrote, and that rule is often the better buy.
- Machine learning needs enough past cases that look like the new case, which is why rare events in a single hotel are usually better handled by a rule.
- Natural language processing and computer vision are fields of application, not separate technologies: either can be built on rules, on machine learning or on deep learning.
- The useful question for any tool is not which technique it uses but what goes in, what comes out, and what happens when the answer is wrong.
Machine learning is software that finds patterns in past records and uses them to score a new case. Deep learning is machine learning with many stacked layers. Generative AI produces new text or images. Natural language processing handles words, computer vision handles images. Each is a tool with a job, not a rung on a ladder.
A proposal lands on your desk with the words "AI-powered" on the cover and three different terms inside: machine learning, deep learning, natural language processing. You have to decide whether the thing is worth the money, and the vocabulary is doing more work than the product description. This page puts the family on one map, with a hotel example for each member and what each needs to work, so you can ask which one the product in front of you actually uses.
How do rules, machine learning and generative AI fit on one map?
They are different tools for different jobs, not stages of maturity. The simplest tool that does the job is usually the right purchase.
AI is the umbrella, not a technology
Artificial intelligence is the umbrella term for software doing work that used to need a person. Under it sit several ways of building. A fixed rule is a condition a person writes: if the stay is longer than seven nights, apply the long-stay rate. The logic is readable, testable, and when it fails you can see exactly where.
Machine learning removes the person from the condition. Nobody writes the rule; the software adjusts numbers inside a model until its answers line up with past records. Deep learning is machine learning built from many stacked layers of simple units, called a neural network, and it needs more examples and more computing than a simple model. Generative AI is a model that produces new text or images; most chat tools people use today are built on deep learning and language models.
Natural language processing and computer vision sit on a different axis. They are fields of application, not technologies: one deals with words, the other with images, and either can be built on rules, on machine learning or on deep learning.
Why the simplest tool usually wins
The commercial point is that the label tells you nothing about the mechanism, and the mechanism decides what the tool can and cannot do. A product can carry the AI label and contain no machine learning at all, and if a fixed rule does the job, the rule is cheaper, readable and easier to change. We cover that trade-off in more depth in when a simple rule does the hotel job better than AI.
| Term | What it does | A hotel example | What it needs to work |
|---|---|---|---|
| Fixed rules | A person writes the condition | Long-stay rate applied automatically | A clear rule and someone to maintain it |
| Machine learning | Finds patterns in past records | Flagging bookings likely to cancel | Enough similar past cases, clean records |
| Deep learning | Machine learning with many stacked layers | Reading a photo of a damaged room | Many labelled examples, computing |
| Natural language processing | Handles words | Sorting guest messages by language and topic | Real guest text, a source to check against |
| Computer vision | Reads images | Scanning an ID document at check-in | Steady lighting, a human check |
| Generative AI | Produces new text or images | Drafting a reply to a guest | A source to check against, a person to send |
What is machine learning, in hotel terms?
Machine learning is software that finds patterns in past records and uses them to score a new case, instead of following conditions a person wrote.
What goes in and what comes out
Past cases go in as a set of facts: arrival date, room type, how far ahead the booking was made, occupancy at the time, the rate finally paid. The outcome goes in beside them. A training step adjusts the model's internal numbers until its answers match those outcomes. The two common shapes are sorting into buckets, such as "will this booking cancel, yes or no", and estimating a number, such as "how many rooms of this type will be left unsold".
For example, a model that flags bookings likely to cancel, so the front desk can call and confirm them. Nothing is cancelled automatically; the flag only decides who gets a call. A close cousin is anomaly detection, which learns what normal looks like for a hotel's own numbers and raises a flag when one moves outside that range, described in what anomaly detection does with a hotel's own numbers.
Machine learning fails in predictable places. A rare event has few examples to learn from, a sudden change in the market makes old patterns misleading, and the model cannot explain a rule it never learned.
How to test a model before you trust it
Hold back the most recent months, run the model on them, and compare its answers with what actually happened and with the simple rule you use today. That comparison is the only honest one.
Watch outA model checked on its own homework
A model checked on the same records it learned from will always look good. Ask for the result on a period the model never saw, and ask what the simple rule scored on the same period.
What is deep learning, and why does it need more data?
Deep learning is machine learning built from many stacked layers of simple units, a neural network, which lets the software build its own features instead of a person naming them.
A plain gloss: a neural network is layers of small number-crunching units. Each layer passes a transformed version of the input to the next, and "deep" simply means many layers. It spread because, for images, sound and language, the useful features are hard to name by hand, and layers can find them from examples.
The cost is more examples, more computing, and less readable reasoning. You get an answer, not an explanation of how it was reached. For example, judging from a photograph whether a room was left damaged at check-out; a simpler tool may do the same job if the question is narrow.
In a hotel, one property rarely holds enough labelled examples to build a deep model from scratch, so most hotels use a model that someone else built on a much larger body of data. When a buyer hears "built on deep learning", that says nothing about whether the output is right for your property. The check on unseen data is what says that.
NoteThe three words are nested
Deep learning is a subset of machine learning, and machine learning is a subset of artificial intelligence. The three words are nested, not competing.
What is natural language processing?
Natural language processing is the field that gets software to handle human language: sorting text, pulling facts out of it, translating, answering questions and writing.
Older systems in this field were rules and word lists. Modern systems are mostly machine learning on large collections of text. The tasks a hotel actually meets include detecting the language of a guest message, translating it, sorting reviews by topic, pulling dates and amounts out of an email, and answering questions about the property.
Words in, words out, and the check in between
Language models are the engine behind the fluent part, and we explain them in why a fluent language model is not the same as a right answer. Fluency is not accuracy, and a confident sentence can carry a wrong room number or a wrong price. For example, a guest writes in three languages asking for a late check-out; the system identifies the request and drafts a reply, and a staff member reads it and sends it.
What it needs: text in the guest's real language, a source the answer can be checked against, and a person for anything that touches money, identity or a promise to a guest. Where it fails: names, room numbers, dates and prices, and any question whose answer is not written down somewhere the tool can reach. The writing side of this field is covered in what generative AI does with words.
What is computer vision?
Computer vision is the field that gets software to read images and video: what is in the picture, where it sits, and what it says.
The mechanism: an image becomes a grid of numbers for brightness and color, and the model learns which arrangements of those numbers matter for the question being asked. Hotel uses include reading an ID document at check-in, scanning invoices and delivery notes, checking a minibar photo, counting people in a lobby, and reading a plate in the car park.
OCR and computer vision are not the same job. OCR turns printed characters into text; computer vision can go further and judge whether the document is the right one or the photo is too blurred to use, a distinction we draw in how OCR reads invoices and delivery notes into hotel systems. What the field needs: steady lighting and a fixed camera angle, a set of labelled examples, and a human check for anything that touches money or identity.
Where it fails: glare, handwriting, a document layout the tool has not seen, or a camera that was moved overnight. One caution before signing: images of guests and their documents are personal data. Ask where the images are processed and how long they are kept.
Which one is inside the product you are being sold?
Ask what the tool takes in, what it gives back, and what happens when the answer is wrong. Those three answers tell you which technology it is, whatever the cover of the proposal says.
If the vendor cannot state the input and the output in one sentence, the mechanism is not clear to them either. Ask what happens on a wrong answer: is there a person in the loop, a threshold, an undo? Ask whether the tool was checked on a period it never learned from, and whether you can see that check. Ask whether a rule would do the same job; if it would, buy the rule. Ask where the data goes, who processes it and how long it is kept. And ask for the status word: operating today, built ahead of launch, in development, or designed. A tool that is designed is not a tool you can rely on this season.
Six questions to put in writing
Checklist
- The input and the output, in one sentence.
- What happens when the answer is wrong, and who confirms it.
- Whether the tool was tested on a period it never learned from.
- Whether a fixed rule would do the same job.
- Where the data is processed and how long it is kept.
- The status word: operating today, built ahead of launch, in development, or designed.
What to do next
Take the next AI proposal you receive and write three lines beside each claim: what goes in, what comes out, and what happens when it is wrong. The map then sorts itself: rules at the base, machine learning and deep learning as pattern finders, natural language processing and computer vision as the senses, generative AI as the writer. Most real products combine several of them, and the ones worth buying can say which.
Frequently asked questions
Is AI the same as machine learning?
No. Artificial intelligence is the umbrella term for software doing work that used to need a person. Machine learning is one way of building it, deep learning is a subset of machine learning, and generative AI is one use of deep learning. A product can be labelled AI and contain no machine learning at all, for example a fixed rule that reroutes a request.
Does generative AI learn from my hotel's conversations?
A model is built once on a large body of text and then stays fixed. It does not update itself from your chats unless someone builds a separate step that stores and reuses them. Looking things up in your own documents during a conversation is a different mechanism and is often confused with learning; ask which of the two a vendor means.
Can machine learning work for a small hotel?
It needs enough past cases that resemble the new case. A small property may hold plenty of ordinary booking records but very few examples of a rare event, and a model cannot learn a pattern from a handful of cases. For rare events a fixed rule is usually the better tool; for frequent, repetitive decisions machine learning can help.
Which of these does a revenue tool use?
Usually machine learning on past bookings and rates, sometimes fixed rules, often both together. The useful questions are which part is learned and which is written by a person, what the tool was tested on, and what happens when it is wrong. A confident label on the proposal is not an answer to any of those.
Is OCR the same as computer vision?
OCR is a narrower job inside the field: it turns printed characters into text. Computer vision covers that and more, such as judging whether a document is the right type or a photo is too blurred to read. For invoices and delivery notes, OCR is often enough and is simpler to check.