To further open up the “black box” of AI, this time a slightly more technical article. I will keep it as clear as possible, even if you do not work with it on a daily basis.
To begin with, it is important to understand that AI is a hype term, and essentially an umbrella term. What exactly does and does not fall under it is not clearly defined. That makes sense, because ICT is a field that evolves every day, and progress moves so quickly that definitions often lag behind.
The AI I am focusing on here is, broadly speaking, probabilistic AI. This is where most misunderstandings arise. It feels exciting, because software used to be fully deterministic. Every line of code was written explicitly, and every outcome was exactly predictable. With probabilistic systems, you work with chances and probabilities, but that does not mean the system is without direction.
From separate tools to one smart interface
Take a chatbot as an example. At Exclusive-IT, we use a self-developed internal AI chatbot every day, if you want to call it that. Not because it is “magical”, but because it is practical tooling.
Think of simple tasks such as converting units, comparing currencies using current exchange rates, or checking the time in another time zone. On their own, these are fairly simple things to program.
The power lies in bringing them together. Instead of ten separate tools each with their own interface, you want one text field where you can simply type your question. That is much more convenient for you as a user.
Understanding human language
And this is where it starts to get interesting. You type into that field the way you speak, in normal human language. The computer therefore has to learn to understand what you mean.
The first step is an NLP, a Natural Language Processor. This is a component that tries to determine the intent of your sentence. In AI terms, this is called the “intent”.
Here is a simple example of training data you could “feed” to the chatbot so it can learn:
{
"intent": "greeting",
"examples": [
"hello",
"hi",
"hey",
"good morning",
"good evening",
"what's up",
"hallo",
"hoi",
"hey daar",
"goedemorgen",
"goedenavond",
"yo"
],
"handler": "smalltalk_greeting"
}With this kind of data, you teach the system how people speak. You link sentences to an intent, in this case a greeting. You also specify which function in the software should handle it, here “smalltalk_greeting”.
That handler is simply a piece of code, written by a programmer, that determines what happens next.
The “magic” is mathematics
To convert this text into something a computer can work with, all kinds of mathematical techniques are used. Think of terms such as tokens, n-grams, embeddings, TF-IDF, and lemmatization.
It may sound complex, but at its core something simple is happening. Words and sentences are converted into numbers and patterns. Based on that, the system can estimate what you most likely mean.
With enough training data, the software can recognize an intent and then call the correct functionality.
From intent to action
After recognizing the intent, there is a next step. The system needs to determine which “tool” will handle the request.
This is often done through a kind of selection process, comparable to multiple options presenting themselves. The tool that best matches the question gets the task. In technical terms, this involves heuristic classification, where lightweight and fast checks are performed first, and only if necessary are heavier calculations used.
This is still not magic. These are simply decision rules and algorithms designed by people.
Facts instead of sentences
When a tool is assigned the task, it does not return a nicely phrased sentence, but raw data. For example, for a currency conversion:
{
"type": "currency_conversion",
"input": {
"amount": 10,
"currency": "EUR"
},
"output": {
"currency": "JPY",
"rate": 186.9
},
"result": 1869
}This response is based on facts that have been collected or are accessible to the chatbot. It is controlled code, written by a human. The goal is very clear and there is no magic involved.
Note that these are purely facts. No sentences, no tone, just data.
From data back to human language
Then comes the post-processor. This component converts the facts into a sentence that is understandable for you. This is often done using templates, which are predefined sentences with variables in them. For example:
“10 euros is currently approximately 1869 Japanese yen.”
Depending on the context, there may be variation. Perhaps “good morning” is added, or the tone becomes slightly more informal. But the facts remain the same.
This is also an important point. The freedom is mainly in how something is said, not in what is being said.
AI is not uncontrollable magic
What I want to show you is that AI is not a mysterious, uncontrollable force. It is a combination of mathematics and software engineering, with clear steps and control points.
Under the hood, it is simply code. Code that is written, tested, and maintained by people. The only areas where there is some “room” are in interpreting language and in formulating responses, but even there, there are strong boundaries.
AI can only go off track if it is designed in a way that allows it to go off track. With craftsmanship, ethical choices, and proper controls, you can build systems that are actually very reliable and useful.
Not magical, but practical. And above all, controllable.