Inside the Machine: How The AI Tools That You Are Using Actually Work

And why every leader deploying them should understand the parts no vendor explains

You’ve approved the budget. You’ve watched the demo. You can describe what it does. But if you can’t explain what happens between the prompt and the answer, you can’t govern it, price it, or know where it will fail you. So let’s open the box to understand exactly how AI tools like ChatGPT and Claude work.

If you’re like most people you can tell someone what the tool does: it drafts, it summarizes, it answers, it writes code. What you probably can’t explain is what happens in the two seconds between the question going in and the answer coming out and neither can most of the people who will now use it for hours every day.

But collectively we have decided that’s fine. We’ve decided AI is one of those technologies you can run a business on without understanding, like electricity or the cloud. But that analogy is wrong, and the wrongness is costing companies real money and real credibility. Electricity behaves the same way every time. These tools don’t. They produce a brilliant competitive analysis on Monday and invent a court case that never existed on Tuesday, and they do both with exactly the same confident tone. If you don’t understand why, you will trust them in precisely the wrong places.

The word in “artificial intelligence” that causes all of the problems is “intelligence.” It’s a smuggler that brings along a whole se of assumptions that we just accept. Things like the system understands you, that it knows things, that it remembers things, that it can tell what’s true and what’s not, and the list goes on. If you take away all of the assumptions you have around AI then you will see and understand the mechanics behind each step.

Intelligence Vs Predicting

Here’s the whole secret behind these AI tools, a large language model (LLM) does one thing, it predicts the next chunk of text. You give it everything up to a point and then it produces a ranked list of what might come next, with a probability attached to each option. Then it picks one, adds it to the end, and does the entire thing again for the chunk after that. And again. Thousands of times, faster than you can read this sentence.

That’s it. The model doesn’t know the final answer when it starts. It works one piece at a time, using patterns it absorbed during training to make its best guess at each step, and keeps going until the response is finished. Everything your organization is betting on in terms of the analysis, the code, the customer replies, etc. is this one move, repeated at enormous speed.

Now if you’re reading this your immediate instinct is probably, “that can’t possible be enough.” Good, hold onto that because it’s the correct instinct to have and I’ll get to it shortly.

Predicting the next word sounds quite trivial but to get genuinely good at it, these models analyze trillions of examples of human writing and content about every possible subject that exists. By doing so, the system is forced to learn the things that make the next word predictable, things like grammar, facts, tone, how an argument is structured, how code is written, and the logic that connects one idea to the next.

The interesting things is that these capabilities aren’t programmed into the AI tools, nobody wrote a rule for grammar or a database of facts. These are just the side effects of getting really really good at a guessing game. These capabilities are of course all real but they are downstream of prediction not the point of it, this is why these tools can simultaneously be world class and catastrophically wrong at the same time.

So to recap, the first thing to remember is that these large language models are not databases that retrieve answers and they aren’t a colleague who understands your questions. They are prediction engines and when you’re confused about why an AI tool did something, come back to that.

Now let’s dive a little deeper into how things actually work and this is where the fun stuff happens…