Left to itself, a language model produces fluent text with no guarantee of truth. Knowledge-grounded AI constrains the model to an authoritative source. Before answering, it consults that source, and its response reflects what the source actually holds.
The benefit is trust. When output traces to defined knowledge, a professional can check it, defend it, and act on it. This is what moves AI out of the "interesting but unverifiable" category and into serious work.
Grounded, not guessed. That is the whole idea.