Building an AI agent yourself

You do not have to program to set up an AI agent. You do have to be able to say exactly what should happen — and that is usually the hardest part, because the work currently sits in someone's head.

Step 1 — choose one task

One. Not a process, not a department. A task that comes back every week, that has fixed steps, and whose result you still want to see before it goes out.

A good first choice: something from administration. You see straight away whether it is right, and the steps are settled.

Step 2 — write down what happens

In plain language, as if you were explaining it to a new colleague. Where it starts, what has to happen, what the result is, and what you do when something is not clear.

That last one is often forgotten and matters most: a task that has an exception once a month needs to know where that exception goes.

Step 3 — point it at the sources

Where it may look: which mailbox, which folder, which system. Only what is needed. An integration goes through your package's official login, with only the permissions it needs, and you withdraw it with one click.

Step 4 — test on work you already know

Set it loose on last month. You already know the outcome, so you see straight away where it differs. If it is not right, you adjust the description and try again.

Step 5 — switch approval on and let it run

For anything that goes out or touches money: approval on. The result is put ready, you say yes, and only then does it go. After a few weeks you know whether it should stay that way.

It does not have to start from nothing

There are 154 ready-made tasks waiting. The quickest route is to take one that sits close to your work and adjust it, rather than starting with a blank sheet.

And if you get stuck, that is exactly what the build session in the training is for: setting up one real task together, on your own figures.