I am putting this distinction in the first chapter of the book because if you do not understand it, you will trust the wrong things.
An LLM does one thing
An LLM, a large language model, does one job: given some text, it works out the likely next word. Then the one after that. That is how an answer is built.
It cannot see anything. It cannot open your container. It cannot get into your GA4 property. All it knows is what answers to this kind of question usually look like, based on what it read during training.
You may have seen the result. Ask ChatGPT "which parameters does the GA4 purchase event need" and it will answer correctly, because that is written in the documentation and it has read the documentation. Now ask "check whether the purchase tag in my container is correct". It will answer, with confidence, and the answer is pure invention, because it has not seen your container.
That second kind of answer is called a hallucination. I see it differently: it is not a hallucination, it is that machine doing its normal job. It was asked a question, it had no data, so it produced a likely answer. The fault is not the machine's, it is the question's.
Agentic AI adds one thing
Agentic AI means the same LLM, but with some tools in its hands, and it can decide by itself when to use which.
A tool is the ability to do something. Read a file. Send an HTTP request. Run a terminal command. And in our case: go to the GTM API and fetch the list of tags in a container.
Now ask the earlier question again, but this time to an AI that has a tool for reading GTM. This time it will run the tool first, get the real list, then read that list and answer.
The difference is night and day. In the first case the answer is a guess. In the second the answer is read from your container, and you can check it tag ID by tag ID.
The whole job of this book is to build that second situation.
But agentic does not mean accurate
Write this down now, because if you forget it later it will cost you.
Having a tool means it will get the right data. Having a tool does not mean it will interpret that data correctly.
I have seen this myself: AI fetched the correct list from GTM, in which one tag's firingTriggerId was empty, and then wrote "this tag fires on every page". The opposite. No trigger means it fires nowhere. The data was right, the interpretation was wrong.
So the rule in this book: AI runs the tools and fetches the data, you read that data and make the decision. And that is why you need to know the tracking work beforehand, as I said in the introduction.
Steps
- Open your AI client and ask: "How many tags are in my GTM container GTM-XXXXXXX?" (put in your own real container ID).
- Read the answer. If it gives you a number, that number is made up, because no MCP server has been installed yet.
- Copy the answer into a file and write the date. After chapter 20.7 you will ask the same question again and compare the two.
- Now ask: "Can you read my container? Which tools do you have available?" Keep this answer too.
What not to do
- Never send a client any number AI gave you without a tool. Ever. If the sentence "Your container has 14 tags, 3 of them broken" did not come from tool output, it is invention, and the client will catch it when they check.
- Do not assume that because AI sounds confident, it knows. There is no relationship between confident language and true information. The worst answers I have seen were the most polished.
Proof
You have two pieces of text, with dates: what AI said about your container without a tool, and what it said when asked to list its own tools. You will need both in chapter 20.7.
Want the full chapter list, or to jump to another chapter?
All chapters in the book