For each part of my business, can it act, propose, or must it ask?
Nobody hands an unfamiliar hire the company card on day one. You put a limit on it, and the limit lives in your finance system, not in the induction pack. Software that writes to clients and moves work along earns the same treatment, and the question above is the shape it takes. Ask it department by department, because the answer that matters is never one answer.
One word, per department
In ArkOne’s reference design for an executive agent, the Cabinet, each specialist works under a standing charter called the Mandate, and every Mandate carries exactly one of three words. Act: an outbound action goes to the guards on what may leave, and if they permit it, it leaves. Propose: the work is finished in full and stops, waiting for a person. Escalate: nothing is drafted, and the matter goes to whoever holds the approval right, with the room’s reasoning attached.
Two things about that word matter more than the three options. It attaches to the specialist rather than the action, so one allowed to act acts on everything inside its charter, and the remedy is a narrower charter. And every specialist ships at propose, a setting on the Register: somebody has to write act, and the writing is dated. The paper on authority takes the three settings a case at a time.
Where the word is kept
Suppose you go department by department in one meeting, and the same vendor answers three times about three parts of your firm.
| What a good answer sounds like | What you will be offered instead | What that hides |
|---|---|---|
| One line per department, one word on each, and a change carries a name and a date | There is human-in-the-loop review before anything goes out | One setting for the whole system. Ask what happens in the department you would trust least, and whether that answer can differ from the rest |
| The default is wait, and here is which departments your people have moved | Approval workflows are fully configurable | Configurable in whose hands. Ask who at your company may change it, and whether the change appears anywhere you can read |
| Authority sits outside the model, so a persuasive request cannot alter it | You can instruct it to always check with you before sending | An instruction the model reads, alongside every other instruction. It is guidance, and this question is about a rule |
Exhibit 1. Illustrative. Three replies about authority, and the permission each of the weaker two grants on your behalf.
The first row is where most buyers stop, satisfied, and it deserves slowing down. A single approval step across everything is a real control and a blunt one. It makes the low-consequence work as slow as the pricing exception, which is how companies come to switch approval off wholesale rather than department by department.
Press the second row, because configurable is true and unhelpful. What you want is who holds the dial, whether a supplier’s engineer can move it during a support call, and whether you would be able to tell afterwards. A control nobody can audit is a control you share with people you have not met.
The follow-up nobody prepares for
Ask what each department shipped with, and who has changed any of them since the pilot began.
The first half costs a vendor something to answer honestly. A product whose departments arrive able to act granted permissions on your behalf before you were in the room, and that nothing has gone wrong yet describes your luck rather than your controls. A product arriving at wait made you choose each loosening, with a date against it.
The second half tells you whether the setting is a control or a display. Ask to see the change history in the demonstration environment, and read who appears in it. Your own people mean the dial is yours. Names from the vendor’s side mean somebody with a support ticket open can widen your agent’s authority faster than your change process can, and that can be acceptable, provided you know it and agreed to it.
Then take the question back on yourself. Which departments would you move to act after a quarter of clean proposals, and on what evidence. A vendor whose answer includes a report you can read has thought about the part after the sale. Set it beside what stands between a decision and a sent message, because act does not switch those rules off.
Authority is the cheapest thing to settle before signature and among the most awkward to correct after. One word, one department, one date, one name. A vendor who can show you that grid has answered half your list already.
component: answer-card
Asked plainly
Can an AI agent be allowed to act in one area and only propose in another?
Yes, where authority is stored as a value against each area rather than described in the instructions given to the model. In the reference design described here each specialist carries exactly one of three words: act on its own, propose and wait for a person, or hand the matter to a person to decide. The routing code reads the word after the model has finished.
What authority level should an AI agent start with?
Waiting for approval, in every area, until a company has evidence it can stop waiting. Starting at wait costs speed and nothing else. Starting at act grants a permission nobody chose, and the first time anybody examines it closely is usually after something has gone out that should not have.
How do I tell whether an AI agent's permissions are real controls?
Ask where a change is recorded. A real setting has a value, a person who last changed it, and a date, all readable without asking the vendor. A described permission has none of those, which is why the honest test is not what the agent does but whether you can point at the line that decided it.