How do you decide when an AI assistant should just act, versus ask you first?

We're a few days out from launching Kurio, an AI assistant that reads your bank account and proposes what to do next rather than just showing you data. Building it forced a decision we didn't expect to be this hard, for every single thing the assistant notices, at what point does it just do the thing versus stopping to ask.

Our answer ended up being a dial per topic rather than one global rule, silent, notify, propose, or auto, so the same assistant can watch your spending silently but always ask before it drafts a cancellation email on your behalf. Curious how other people building agents here have thought about this, especially anyone working on something that touches consequential stuff like money, calendars, or contracts. Where did the line end up for you, and did users push back on the defaults you picked?