Sarah asks a firm's AI chat four reasonable questions about her pension and gets four correct answers. Nothing it tells her is untrue. An adviser calls, twice, and does everything right. She had already decided three minutes earlier, in a chat that told her neither of the two things that mattered.