The problem
A language model understands quickly, but it can misunderstand. It can get the recipient, the date or the tone wrong. When it only answers, the error costs you a reread. When it acts, by sending an email, moving an appointment or deleting a contact, it can cost much more.
On the other hand, an assistant that asks permission for everything quickly becomes tiresome. If every calendar read triggers a question, you end up approving without reading, and the protection is useless.
What I wanted
I wanted LIA to ask for your approval at the moment the action is being prepared, on what is really going to go out, and only when it's useful. Reading your calendar or searching for an email asks for nothing. Writing, sending, changing or deleting goes through you, apart from a few actions with no consequence for anyone else that can be undone in one click, like checking off a task or turning on a lamp.
How it works
You say: “move my Thursday appointment to Friday”. LIA prepares the change and shows it to you as a card: the event, the before, the after. You approve, you refuse, or you reply “make it 3 p.m.”. The card has its buttons, but you can also just answer, in writing or by voice.
- One draft at a time. If your request produces several, for example two emails and an event, LIA presents them one by one and tells you where you are (“draft 2 of 3”). Nothing goes out before your last answer.
- What you approve is what runs. If you ask for three tweaks, it's the last version shown that goes out, and nothing else. If the resource has changed in the meantime, LIA shows you the preview again, since your approval was for the old version.
- An honest report. Afterwards, LIA separates what succeeded, what's partial, what's waiting and what failed. It doesn't announce as done what a service refused.
When nobody is at the screen
LIA also works without you: scheduled routines, tickets it takes on. There, two rules apply. On a ticket, if it prepares something it isn't allowed to do alone, it asks you on the ticket, with the card you would have seen in the chat; you answer with a comment. A routine, on the other hand, refuses the action and says so, because otherwise your next message, written about something else entirely, could pass for an answer.
The exceptions, and why
Some actions can't reach anyone but you. LIA can send an email to your own address, or call you on a number you've verified, without a confirmation card, including from a routine. Likewise, keeping, editing or discarding a draft in your mailbox asks for nothing; it's sending it that asks for your approval, just once. An address you type yourself, even your own, still goes through the usual confirmation: a recipient that sometimes skipped the check would be too easy a target.
For MCP, you choose confirmations at account, server and tool level. What it will do explains the effective rule before you change it. A dedicated agent asks before its task when one of its tools requires approval; that answer covers its calls for the task. A tool’s “never ask” choice is suspended when its declaration changes. The chat may strengthen these protections; only Settings can reduce them. A routine may use the servers you explicitly pre-approved in its form.
The limits
A confirmation covers the action shown, not everything an external service will do afterwards: when LIA delegates to a tool, it can't prove each of that tool's internal effects. And when a request requires connecting an account, entering a key or choosing a file, LIA prepares it, then lets you finish in your browser, where your identity is verified.
The principle fits in one sentence: LIA prepares, you decide, and then it tells you what really happened.
Read more: Protecting your account and your data: LIA's security, layer by layer; When the model makes things up: how LIA guards against it.
