When Is an AI Answer Actually Grounded?

A team asks an AI assistant to answer from internal documents. It cites three relevant pages, but one key recommendation is an inference that none of them states directly.

What should count as enough grounding to use that answer in a work decision? Is it enough that the sources are relevant, or should the team also check whether each important claim is supported—and whether the assistant marks its own inferences clearly? Would your threshold change with the consequences of getting it wrong?