Why you review AI
FIA suggests drafts and explanations. The book, not the suggestion, is where saved amounts and balances come from.
A receipt is evidence, not an instruction. A model that reads "total 12.50" can still pick the tax line, the wrong day, or a currency from a dollar sign. A statement can include yesterday's balance, which is not a purchase. If that figure saved itself, the book would be whatever the suggestion guessed, and a budget would inherit the guess.
So FIA produces a draft. You see the field next to the words or the page it came from. You correct it, or you leave the row out. Save writes the same kind of entry a typed one would. AI never posts to your book on its own. Discard asks Discard what FIA read? and says the entries waiting there will not be saved. Confirming it writes nothing. A second save of the same file does not write it again, which is the other half of the same rule: review is not a slot machine you can hit twice.
The assistant is the same idea in the other direction. It may only talk about records you can already open. The app calculates the totals. The model is not allowed to invent a balance, and a sentence that states a figure the report does not support is dropped. A suggestion to change a budget is a proposal, approved like any other change, once. "If we save 200 a month" is a scenario you are reading, not a transfer it performed.
That is also why there is no remote Chicken Bank model. If Apple's on-device model is unavailable, you still have typing, and whatever the document already makes readable. A rule you authorise, such as recording a bill on its day, is not the model deciding. It is a version of a rule you allowed, and it waits when the case is one you did not allow.
Last updated on