When AI stops answering and starts doing

Until recently, using AI was like asking that colleague who knows everything, over coffee: you ask, they answer, and the work of doing something with the answer is still yours.

That is changing. A new wave of tools, called AI agents, does not just answer: it takes a whole task, plans the steps, edits files, researches, writes and delivers the finished result. Several products already do this, each with a fancy name. The principle is the same in all of them.

Think of an intern

The best way to understand it is to imagine a new intern who just joined your team. They are absurdly fast, never complain, work overnight and know a bit of everything.

But they are an intern. They:

• Do not know your company, your family, your routine. They only know what you tell them.
• Do not say "I don't know". When in doubt, they invent a plausible solution (remember the edition on hallucination?).
• Have no sense of what is serious. Sending the wrong e-mail to a client and getting the color of a slide wrong weigh the same to them.

Nobody hands the intern the keys to the safe on day one. With AI, the rule is the same.

What changes in practice

Chat was a conversation. An agent is a delegation. And delegating well is a skill, not a button.

1. Describe the whole task, with the expected result. "Summarize this report" is vague. "Write a half-page summary, in plain language, with three recommended actions at the end" is a task.
2. Start with what repeats. That Monday report, the spreadsheet you build every month, the meeting summary you always write. Repetitive work is where the intern shines and mistakes cost little.
3. Give access only to what is needed. If the task is to read folder X, it does not need the whole drive, nor your bank password.
4. Review before anything goes out. No e-mail sent, post published or money moved without your approval. Serious tools ask for that approval. If yours does not, be suspicious.
5. Ask for the log. A good agent tells you what it did, step by step. If you cannot know what it did, you cannot trust it.

Does this take my job?

It takes part of it: the part you already did on autopilot. What remains is precisely what requires judgment: deciding, noticing that something is off, knowing what matters to that person, taking responsibility for the result.

People who work in AI governance call this "human in the loop". In plain words: the machine prepares, the person approves. It is the seat belt of this story.

Keep this

An AI agent is a brilliant intern with no sense of risk. It does, you answer. So review.

Exercise of the week: pick a task you repeat every week and write, in three sentences, how you would explain it to someone on their first day at work. That explanation is your first request to an agent. And, as a bonus, it shows how much of your work can be delegated without losing control.