The hardest problem was the assistant. A rabbi doesn’t think in records — he thinks “Carl Feldman handed me a check at the bris today, for $180, write him a tax receipt and remind me to check in every two weeks.” That one sentence is three database writes, a document and a recurring reminder. Parsing it was the easy half. The design problem was trust: an assistant that silently writes to your books is a liability, not a feature. So the reply isn’t a paragraph — it’s a stack of confirmation cards. Each card carries one action, the exact amount, the check number and the receipt flag, and nothing is committed until it’s confirmed. The copy says so out loud. The assistant is allowed to be confident and still be wrong, because the user holds the last click. That single rule shaped every AI surface in the product, and it’s the reason the feature reads as an assistant instead of an autopilot.