Observe
Read the situation and keep the picture current. No output, no side effects.
Guide
AI That Asks Before It Acts: Why Approval Boundaries Matter
The important question about an assistant is not how much it can do. It is what it will do without asking you.
An assistant that asks before acting separates what it knows from what it is permitted to do. Guardian Daily is built on that separation: it can understand a situation and prepare a response, but consequential actions wait for a person to approve them.
Opens the Guardian Daily app. You may need to sign in or create an account.
The principle
Knowing is not permission
Most failures of automated systems are not failures of understanding. They are actions taken correctly on an incorrect assumption.
- An assistant can be right about the facts and wrong about what you wanted.
- Context it cannot see — a conversation, a mood, a history — often decides the right answer.
- Reversing an action that other people already saw is rarely possible.
- So the authority to act should be granted separately from the ability to reason.
In practice
Four levels of what an assistant may do
Suggest
Surface a conflict or an option. You can ignore it entirely.
Prepare
Draft the message, build the reschedule, assemble the booking — but do not send it.
Act within limits
Carry out only what you explicitly permitted, and record what was done.
How to evaluate a tool
Questions that reveal real boundaries
- Can you see a list of everything it is allowed to do without asking?
- Can you narrow that list, not only widen it?
- Is there a record of every action it took, and when?
- When it is uncertain, does it stop — or does it pick the most likely option?
- Can you revoke a permission and see the effect immediately?
Why it matters at home
Consequences land on people, not systems
Other people's time
Moving a meeting or cancelling a plan spends someone else's day as well as yours.
Relationships
A message sent in your name is read as you, whatever generated it.
Responsibility
Accountability cannot be delegated to software. Conscience remains human.
Questions
Common questions
- Isn't asking for approval slower?
- For trivial actions, slightly. For consequential ones, it is the difference between a suggestion you can ignore and a mistake you have to undo in public.
- What counts as a consequential action?
- Anything another person sees or that is hard to reverse: sending a message, moving someone else's time, spending money, sharing information, cancelling a commitment.
- Can I give it more freedom over time?
- Yes. Limits should be something you widen deliberately for specific categories, not a default you have to discover after something goes wrong.
- What should it do when it is unsure?
- Stop and ask. Confidence is not permission, and knowing something is not the same as being allowed to act on it.
Keep reading
Related guides
More plain-language explanations of how Guardian Daily is designed.
Human-in-the-loop assistants
What the phrase means in practice.
Privacy-first personal AI
Memory, deletion and export, explained.
Using AI as a daily planner
A realistic plan you still control.
AI assists. You decide.
Guardian Daily prepares the work and waits for your approval on anything that carries a consequence.