An instruction can be misunderstood.
“Don’t spend more than $500” is useful guidance. But if the only place that rule exists is in the AI’s instructions, you are relying on the AI to remember it, interpret it, and obey it every time.
A card with a $500 limit works differently. The card provider checks the transaction. The AI can be convinced a purchase is necessary and still be unable to make it.
We build Busy around that distinction. Instructions explain the job. Permissions decide which actions are available.
The model and the controls have different jobs.
Model makers train AI to refuse harmful instructions and resist manipulation. Busy benefits from that work. It matters when a Persona reads an email, visits a website, or encounters instructions that conflict with its job.
We also know the company it works for, the role it has, and the systems it uses. That lets us set specific boundaries outside the model: who it may contact, what an account can change, and when approval is required.
Those controls need to hold on the day the model gets something wrong. A confident explanation doesn’t grant a permission.
There is more than one place to enforce a limit.
Your existing software enforces account permissions. A read-only role can’t edit the books. A shared folder gives access to that folder. Your card provider enforces its spending controls.
Busy adds checks around actions, including recipient rules and approval requirements. The check runs before the action goes through. If a required check can’t complete, the action stops.
The Persona runs on a computer it doesn’t administer. It cannot rewrite the controls governing its work. Changes to those systems go through reviewed code and automated checks.
Limits make delegation practical.
The goal is to give Busy work it can finish without asking about every step. A bounded card lets it make routine purchases. Read-only access lets it review the books throughout the day.
You still set the job, teach it your business, and review work that needs judgment. The controls put a boundary around what it can do while it learns.