How we decide what Leila is allowed to do alone
The line between what an assistant should just handle and what has to wait for you — and why we drew it where we did.
Every useful assistant has a line. On one side is what it just does; on the other is what it brings to you. Get the line wrong in one direction and you have a product that asks permission for everything, which is a to-do list with extra steps. Get it wrong in the other and you have something that sent a message in your name that you would not have sent.
We have moved that line exactly once, and we expect to move it rarely. Here is the rule we use.
The test: can it be taken back?
If Leila gets something wrong and the consequence is that you read a slightly worse summary, that is recoverable. If the consequence is that a payment left your account, a form was submitted with the wrong answer, or a teacher received a message in your voice that you would not have written, it is not.
So the rule is not about how confident she is, or how routine the task looks. It is about whether an error is reversible. Confidence is the wrong test because the cases where a model is most confidently wrong are exactly the cases you would most want to catch.
What she does on her own
Reads every approved sender as it arrives and summarises it. Works out what actually needs doing and what is a notice. Fills a consent form from what she already knows about that child, including the allergy you would have had to remember. Drafts the reply in your voice. Puts the dates and a sensible reminder in your calendar. Briefs the nanny and the driver on their part of the day, and no more than their part.
All of that is preparation. None of it leaves the house.
What always waits
A payment. A signature or a consent. A message sent in your name to anyone outside the household. A change to a child's medical or dietary record. Adding a new sender, or adding a new person to the household.
Those five are in the terms of service, not just in a help page, and we have committed that if we ever want to move the line we will ask first and you will be free to say no and keep using the product. That is a deliberately awkward commitment to make, which is rather the point of making it.
Why not just add a confidence threshold?
Because thresholds drift. Every product that has ever shipped an 'auto-approve when we are sure' setting has, under pressure to feel magical, become sure more often. A categorical rule about reversibility cannot drift, because it does not depend on a number anybody can tune.
It also makes Leila easier to trust in the only way that counts: you can state, in one sentence, what she will never do without you. If you cannot say that about an assistant that reads your family's mail, it does not matter how good the summaries are.