01
Put permissions before intelligence
Financial providers are connected through read-only access. The product has no payment initiation path, and the language model never receives bank credentials or provider tokens.
A separate security boundary prevents a convenient assistant feature from expanding into authority over money.
02
Keep numbers deterministic
Cash coverage, protected buffers, known bills and category limits are calculated by code with inspectable rules. Missing data reduces confidence instead of inviting the model to fill the gap.
The same inputs always produce the same numeric state, which makes testing and review possible.
03
Give the model a bounded job
The model receives sanitized aggregates, confidence and allowed coaching context. Its task is to explain trade-offs, ask for missing information or reframe the next action.
It cannot move money, change the ledger or redefine a safety rule.
04
Design for incomplete signals
Open-banking data can be delayed or rate-limited. The interface must show last-sync time, coverage and provisional status instead of presenting stale data as real time.
Safe AI depends as much on honest uncertainty as it does on prompt quality.