The personnel also need to take the fall when the AI does it wrong. A judge isn't going to jail Claude, they are going to jail the sucker who unknowingly authorized the fraud.

Will Anthropic externalize the risk, selling access to agents? Or will internalize the risk and liability, selling financial services? Maybe both? I guess lots of companies want both, doing some things internally and keeping other things at arms length by outsourcing to 3rd party accountants.