From Prediction to Action: Why Probabilistic AI Needs Deterministic Boundaries

Andrés Lage Freire starts from a simple idea: an AI model can propose, but it should not govern an action with real consequences on its own. Language models are probabilistic by design and that uncertainty cannot be removed, so the real risk appears when an answer stops being text and becomes an action that changes data, triggers a transaction or is delegated to other agents.To contain it, the talk proposes a deterministic envelope around the model, with two levers, reducing the probability of error through validation, adversarial testing and monitoring, and reducing its impact through least privilege, caps, human approval and rollback, so that delegating capability never increases authority. It frames this as a separation of powers inspired by Montesquieu, in which a legislative layer sets policies and limits, the agent executes within the permitted scope and an independent judicial layer verifies, records evidence and can block or stop the action, because an agent should not be the sole judge of its own actions.The talk closes with a two-act demonstration of the HR Copilot, the same model and the same task with and without runtime control, and with the loop between governance and red teaming, in which every adversarial finding becomes a deterministic rule and a regression test. The ideas in the talk are developed in the white paper Probabilistic Models Require Deterministic Governance, available in Spanish on this page.