What an AI secretary actually does (and what it should never do alone)

Almost everything sold today as an "AI assistant" is really a search box with good manners. You type, it answers, and the relationship ends there. A secretary, in the old sense of the word, does something rather harder to build: it looks at your day before you do and tells you what is going to hurt.

Reading is not the hard part

Summarising a long email has been a solved problem for years. What is still not solved is deciding which email deserves to interrupt you. A model can rank a hundred messages by importance in a second, but importance is not in the text: it is in your week. The same message from a supplier is noise in January and an emergency the day before a delivery.

That is why an assistant that only reads mail falls short. It also needs to see the calendar, the tasks and the commitments you already made, and cross them. It is the difference between "you have 40 unread emails" and "of those 40, three affect something you promised for Thursday".

The three things it should never do alone

The more useful an assistant becomes, the more dangerous it is to give it a free hand. Three lines are worth not crossing without explicit confirmation:

• Sending on your behalf. Drafting, yes. Pressing send before you have read it, no. A badly sent email cannot be recalled, and it carries your name, not the model's.
• Bulk deleting or archiving. Archiving is reversible; deleting, depending on where, is not. An assistant that "tidies" your inbox with no record leaves you unable to find out what you lost.
• Sending data outwards. Any call to an external service takes your information out of the system. That should be your decision, not an implementation detail.

One rule covers all three: if an action is hard to undo or leaves your control, ask. If it is reversible and stays inside, do it and say so.

Autonomy in steps, not as a switch

The usual design mistake is to treat autonomy as a yes or a no. Steps work far better: first it only watches, then it proposes, then it performs the reversible things, and only at the end — if you ask for it — the irreversible ones. Each step is earned with a track record, not with a checkbox in settings.

Above all, you should be able to answer three questions at any moment. What does it know about me. Why is it flagging this. What exactly happens if I press this button. If any of the three has no clear answer, the assistant is not ready for more autonomy, however well it summarises.

You can see how we approach it in the Madrugo console, and what data it keeps and for how long in our privacy policy. If the subject interests you, there are more articles on AI assistants.