Is it safe to put sensitive client information like Social Security numbers into AI tools like Claude or ChatGPT?

The group flagged this as a key open question for the working group to address, since being able to safely use sensitive data would allow much more automation of accounting workflows. One participant's initial take was that the mix of 'scripting versus putting stuff in an LLM' is something they're still working out. Even though they use coding agents for almost everything now, they found that the majority of what they build is scripting rather than sending sensitive information into an LLM query directly. Their broader point was that, unlike some other industries where data sensitivity dominates the conversation, many of the most useful things accountants can do with AI may not actually require putting sensitive information into an LLM at all. The discussion was cut off before a full example was given (an example involving a month-end reconciliation task was starting to be described when the segment ends), so this remains an open, unresolved discussion.