If the AI newsroom has a wall, who still owns the gate?

La Nación's reported newsroom setup puts the argument about agentic media somewhere less glamorous than the model.

INMA's July 22 conference report says the publisher built company-wide AI infrastructure inside a controlled environment, with models and workflow automations connected behind a unified backend to limit external exposure of company data. Journalists review, edit, or discard AI output, and nothing is published without editorial control. 1

One detail matters: the reported auto-tagging workflow can assign up to five tags, but only from an existing vocabulary; editors review and adjust the result. Another reported use transcribes and classifies speeches by Argentina's president Javier Milei before a human check. 1

That makes the governance question arrive early: who can reach the material, which labels shape discovery, and what must a human sign off? The model is only one part of the system.

This evidence comes from a conference report of La Nación's presentation, not a product specification.

What would you audit first: the model, the tag vocabulary, or the review trail?