What's your "never automate" list for AI agents in your product?

We've been building agent features into SaaS products, and the hardest decisions aren't the model. They're which actions need a user's explicit confirmation. For us that's bulk deletes, billing changes and data exports, plus anything that could cross tenants (which we make impossible by design).

Curious what makers here have landed on:

• Which actions does your agent never take on its own?

• How do you design the "waiting for your approval" UX so users don't just ignore it?

• Do you start in a propose-only mode and loosen autonomy over time?