Rex：专注订单到现金流程的 AI Agent 讨论

订单到现金流程是 AI Agent 一个极佳的切入点——量大、重判断，而你描述的“每天从门户复制到 NetSuite”这种枯燥工作，恰恰是自动化历史上经常翻车的地方。我接受你“请问他们哪里做得不对”的邀请，@merlin_kafka：

Agent 如何决定自主执行与转交人工之间的边界——是采用置信度阈值、按操作类型设定的固定策略，还是基于学习的动态判断？

具体到门户上传功能——我做过大量针对网站的浏览器自动化，那些网站经常毫无征兆地变化，这正是最容易变得脆弱的环节。你们是如何在上百个客户门户中保持这些流程稳定的？

另外，由于错误的应收账款操作会直接涉及真金白银，你们在发出指令之前，如何评估 Agent 确实做了正确的事情？

Order-to-cash is a great wedge for agents — high-volume, judgment-heavy, and the "copy from portal → NetSuite every single day" drudgery you described is exactly where automation historically fell over. Taking you up on the "ask what they get wrong" invite @merlin_kafka:

How do the agents decide the boundary between acting and routing to a human — a confidence threshold, a fixed policy per action type, or something learned?

On portal uploads specifically — I do a lot of browser automation against sites that change without warning, and that's usually where things get brittle. How do you keep those flows stable across hundreds of customer portals?

And since a wrong AR action touches real money, how do you eval that an agent did the right thing before it goes out?