You have probably used AI in a chat window. You type something, it answers, and it stops. Then you type the next thing. Whatever the job is, you are the one moving it forward, one message at a time. This is Chat AI.
Agentic AI is different. An agent is the same AI ‘engine’ that you use in Chat AI with two things added: tools it can use to do real work and the freedom to keep using them until the job is finished. You hand over the whole task rather than the first step of it, and it works through the task without you.
That is the difference, and it sounds like a small one. It is not. Everything else in this course follows from what happens when nobody is watching each answer arrive.
To see why it matters that much, it helps to know what kind of thing you are handing the job to.
Almost all of the software you have ever used, whether on your computer or on your phone, has until recently worked the same way. Someone wrote the rules it follows, and it follows them exactly. Add up the same column of figures twice and you get the same total twice. Check your calendar for today’s events and it displays the events for that day. That reliability and predictability is so ordinary that you have probably never once thought about it.
AI is a different kind of software. Nobody could write down a rule for every question it might be asked, so it was built another way: it was trained on an enormous body of text (that’s the ‘L’ in Large Language Model) and examples until it could work out what a good answer looks like. Ask it something and it produces the most plausible answer it can.
So what you get back is probably right.
For a great deal of work, probably right is what you want. You ask for a draft, an explanation, a second opinion on something you already understand and you judge the answer as it arrives just the way you would judge a suggestion from a colleague. That is the chat window, and it is most likely how you have used AI so far. And generally AI gives you an answer which is probably useful. But it’s up to you to check that what it’s telling you is actually true.
‘Probably right’ is a different proposition when nobody is reading and checking the answers. Ask AI to run the same job a hundred times when you are relying on the numbers that come out and you need better than plausible. You need a correct answer, every time. You don’t want ‘Probably right’. You want a correct, reliable answer.
The way professionals ensure that they get accurate answers is to build a framework around the AI: a set of checks and limits that behave the same way every time whatever the AI does on any given run.
This course teaches you how to build that framework. The AI at the centre of it stays exactly as it is, and no instruction can ever make it truly reliable. What does make it reliable is the framework around it. The industry term for an AI working inside a framework like this is an agent.
To illustrate this we’ll use an example. Picture, for a second, a job you routinely put off until the end of the month. A folder filled up with receipts, invoices, statements, a stack of case notes, whatever your version of the paperwork is. On the last Friday of the month you sit down and work through the stack by hand, one item at a time.
You have probably already tried to get AI to help with a task like this. You open a chat window and paste in the first receipt. The assistant reads it back, pulls out the date and the amount, and waits. So you paste the next one and the next one. An hour later you are still sitting there, feeding the items in by hand and copying each answer back out into the spreadsheet. Every individual answer arrived in seconds. It still takes the whole afternoon.
What you are using is AI as a chat assistant. When you use Chat AI it answers the question or message you give it and then stops. It moves on only when you type the next thing. You are the engine that keeps it going. And it’s up to you to check that what it has given you is accurate.
What changes when you hand over the whole folder instead to an Agent
An agent is an AI that has been given tools it can use to carry out a task and the freedom to keep using them until the task is finished. You will build one during this course, and you will do it by talking to it in plain English, exactly as you already talk to AI now.
The difference is which tool you do that talking in. The chat window you have been using is designed for small amounts of information that you hand over in the conversation itself: a document, a question, a draft to look at. It does that well. It is not built for a folder of two hundred items.
Claude Code is designed for just this kind of work. It is an application from Anthropic, the company that makes Claude, and it runs on your own computer rather than in a browser tab. You point it at a folder and it works on the folder, including the documents you add to it next month. That is what it was designed to do, and it is why you will use it for this job. There are other AI Agents as well, such as Codex from OpenAI, but for this course we are using Claude Code. The prerequisites page lists what you need and how to install it.
Although tools like Claude Code and Codex were built for programmers, they are very useful for anyone who needs to work with large amounts of data in a reliable fashion. Working through a folder of real work was never only a programmer’s problem, and you do not need to be one to use their tools. You will type the same plain English you would say to a colleague.
With Agentic AI you do not hand the AI one item and wait. You hand it the whole folder, together with your instructions, and you let it run. In plain words, something like: go through every document in this folder, pull out the date, the supplier and the amount, put each one on a row in this sheet, and tell me which ones you could not read clearly. Then you step back.
The Agent is built to work through the pile on its own rather than returning after the first receipt to ask what happens next. Most of the time it will keep going until the folder is empty, though it may pause to check something with you, stop earlier than you expect, or hand back a receipt it cannot confidently handle. Keeping the work moving used to be your job. Now it is the agent’s, and your afternoon is your own.
What you give up when you stop watching
That sounds like a convenience, a way to save a few clicks. It changes something larger than that: what you are responsible for.
With the chat assistant you saw every answer as it appeared, but seeing an answer is not the same as checking it. A wrong reply that looks right can slip past you, so watching the work does not on its own guarantee the mistake is caught. You were the check, whether you thought of it that way or not. An agent working through two hundred documents while you make a cup of tea has no such check. If it misreads a date on item ninety, nothing stops it, and it might not mention it afterwards. The spreadsheet you come back to looks exactly like the one you wanted.
The power Agentic AI gives you and the risks of it making mistakes are two sides of the same coin. On the one hand it automates work that otherwise would require manual processes which can save both time and money. On the other hand it can make mistakes, sometimes serious ones. You cannot have one without the other.
The solution is not to try to fix the AI. You can’t. The model is what it is and it is not something that the end user (you) can change. What you can change is what it is allowed to do. Before you can do that, though, you need to know precisely what you are working with: what this new colleague brings to the job, and what it will never bring unless you supply it. That is where the next lesson starts.