Home Assistant’s new guide to the existing AI Evaluate action shows how an automation can ask questions about text, structured data or a camera snapshot and use the answers in its next step. Questions can request an estimated yes probability, a choice between named alternatives or a score against an ordered set of criteria.

The worked example evaluates a front-door snapshot when the doorbell rings, then thanks a delivery driver through the doorbell speaker if the model’s estimated probability reaches 90%. That number is the model’s estimate, not a guarantee. The example needs a camera, an AI task entity that supports evaluation, and a doorbell capable of the required text-to-speech playback. The guide covers both the automation interface and YAML; the AI entity must support evaluation, which is separate from data generation.