In theory the whole process is completely automated, just with the ability to manually rewind if I disagree with some of the decisions. In practice, I'm always tweaking the workflow and trying new things that I'm not ready to code up yet.

^redacted‌ I’ve seen interest in this growing, but most of the practical use I’ve come across so far is assistive rather than fully automated. Local models can be useful for things like: reviewing documents without sending sensitive data off-platform spotting inconsistencies or patterns across CIMs, QoE drafts, contracts, etc. generating questions to investigate, rather than answers to rely on Where I’ve seen people be more cautious is fully automating judgment-heavy areas (tax, legal, deal structure). The risk isn’t just that the model can be off — it’s that it sounds confident enough that you stop interrogating the output.