If she never uses ai at work at all, start with something dead simple: -do research with it -put a large corpus of documents into LLM context and ask questions of it
That's probably where I'd start.
Honestly that wouldn’t be a terrible product: Pull in all of a judges written judgements or stated logic about their rulings and preemptively craft your case around their biases.
Maintaining privilege depends on taking steps to ensure the continued confidentiality of the data - put a lock on your file cabinet, sign an agreement with the vendor where they promise not to look at or use the data for their own purpose.
Even if hacker news thinks that agreement isn't ironclad doesn't mean privilege is or should be destroyed. It's more about whether you limit the distribution of the data in a reasonable way. Today, zdr remains reasonable.