What am I announcing, to whom, and in what order?

You describe the announcement. The assistant proposes segments — a subject, a set of countries, a medium, a prominence band — and prices each one against the live corpus before you accept it. Accepting a segment turns it into a saved list you can edit by hand.

Because the model chooses filters and the database returns people. Every count it quotes comes from the same function the campaign composer uses at send time, so a planning number and a send number are the same arithmetic and can be compared.

It is given counts, vocabulary and outlet identity — never a journalist’s name, role or address. That boundary is described in full on the AI transparency page.

It is handed the real vocabulary with live counts, in the languages the corpus actually uses — Cronaca Locale, Wirtschaft, Economia — including the zero-count entries, so it can tell “no such subject” from “nothing classified yet”. That distinction is the one job here a query cannot do and a model can.

Send anything. No tool available to it writes a recipient, moves a campaign out of draft, verifies a sender or touches the suppression list. Where arithmetic beats a model, no model is called at all — narrowing an audience in the composer is four counting queries and nothing else.