We don’t prompt our
way to safety.
We put the limits in the software. Your Persona can work within them, but it can’t change them.
A rule should hold even when the AI gets it wrong.
“Please don’t spend more than $500” is an instruction. A card that stops at $500 is a control. Busy uses controls.
We set permissions outside the model: what it can access, who it can contact, and which actions need your approval. The AI’s confidence doesn’t change those permissions.
Set by you. Enforced by the system.
Its own accounts.
Only the access it needs.
Your Persona signs in as itself. Share a folder, give it a read-only role, or let it use a particular part of your software. Those account permissions apply just as they would to a colleague.
You can see who did the work in the activity log. To remove access, unshare the folder or suspend the account.
If a job needs access to your own inbox, that’s a separate, explicit grant with its own limits.
When it needs permission,
you get the decision.
Busy pauses the action and sends an approval request to your phone. You see what will change before you decide. We call this Two-Factor Override, or 2FO.
Approval happens outside the Persona’s control. It releases the action you reviewed. If you decline, or the request expires, the action stays blocked.
Raise “rice shampoo bar” from $0.52 to $0.57.
You decide who it can talk to.
Keep it working with you alone, open it to your team, or allow contact with people in an approved list. Wider access is a choice you make for that Persona and its job.
When you use an approved list, the Persona can’t give itself permission by adding someone to it. Adding a new recipient needs a separate approval.
“The document is ready and I’ve drafted the email. Dan isn’t on my approved list, so I’ve sent you an approval request. Nothing has been sent to him.”
If the check fails,
the action stops.
An unavailable permission service doesn’t grant access. An empty recipient list allows nobody. An unanswered approval request doesn’t turn into a yes.
Busy tells you what stopped, why, and what’s still waiting. The draft or unfinished work remains available.
Can’t check the recipient?Don’t send.
Can’t confirm permission?Don’t proceed.
No answer to the request?Keep it blocked.
Keep one mistake from becoming a thousand.
AI can misunderstand a request. The damage depends on what it’s allowed to do next. A refund for one customer and a refund for every customer are very different permissions.
We limit access, spending, and the scope of an approval so one wrong decision has less room to spread. Model training helps too. It’s one layer of protection alongside account permissions, Busy’s controls, and a dedicated computer the Persona doesn’t administer.
A closer look at safety.
Five short reads on how the controls work.
- 01Rules the AI can’t changeWhy instructions need to be backed by permissions.
- 02Its own accounts. The right access.Identity, account permissions, and a computer of its own.
- 03Who it can talk toFive settings for replies, outside senders, and new conversations.
- 04The decision comes to youHow 2FO approves an exact action outside the Persona’s control.
- 05When things go wrongFailed checks, unanswered requests, and keeping mistakes contained.
Clear limits.
More room to work.
A Persona with a spending cap can handle routine purchases. One with read-only access can work through the books. You decide where it needs you, so you don’t have to supervise everything else.
Get Busy