Feature lists converge; architectures don't. The question that actually predicts behavior is whether the agent reads and writes the same records your team does, in real time, or operates on a copy and reconciles later.
That single decision cascades into everything: how fresh the data is, how much context the agent has, and whether "the AI did it" and "the CRM shows it" are the same moment or two different ones. Architecture is what you're really choosing.