ARKONE
← The Cabinet
Case note · silentThe Signature · the human approval step5 minute read

The agent that signed as the CEO

Whose name an agent signs and whose account it sends from are two settings, and neither belongs to the model.

The Signature, seen from above on the Cabinet floor plan.
Exhibit 1. The Signature, at its place in the room.

Silent. A supplier receives a short message about renewal terms. It is polite, it is accurate, and it ends with the chief executive’s name above the chief executive’s title. It arrived from the chief executive’s own address. She has never read it. She will not learn it exists unless the supplier replies and thanks her for it.

Nothing in the system registered an event. If you are the chief executive, this reaches you as ordinary correspondence written by somebody else, and to everyone on the outside it is your word.

What went wrong

Two separate settings decide who a message appears to be from, and neither of them is the model’s to choose.

The first is the account. Which mailbox, which handle, which number the message leaves through. That is configuration in the sending system, set once by whoever wired the agent up, and easy to point at a real person, because a real person’s account is the one already working.

The second is the name at the bottom. That usually lives in the drafting instructions, as a line telling the agent whose voice to write in. Give it the chief executive’s name and it will sign as her, faithfully, every time, because signing as instructed is the whole of its job.

Nothing in that arrangement joins the two settings to each other, or either of them to the question of who read the message. So you can arrive, by two reasonable decisions taken weeks apart, at a message nobody approved going out under a signature that means something.

The Cabinet, ArkOne’s reference design for an executive agent, keeps these three questions apart on purpose. The Signature, the step that waits for a person, holds an action until a named human decides, and when they answer it records the decision and stops rather than carrying on. The Mandate, the setting for what each specialist may do unasked, ships at propose and wait, so acting alone is something a person writes in and the change is on the record. And the Directory, the roster of the real people the room may address, names approval rights per person: exactly one scope covers legal signature, and it belongs to the chief executive. The seventeenth paper sets out the waiting step; the twentieth sets out the roster, and both settings are on the Register.

The record

Suppose the distribution firm below, its people and its correspondence, all invented for the exercise, including the firm.

The message The setting that decided it Who approved that setting
Left from the chief executive’s mailbox Sending account, in the mail configuration The person who connected the mailbox, in week one
Signed with the chief executive’s name and title A line in the drafting instructions naming the voice The person who wrote the instructions, in week three
Said the renewal terms would hold The agent’s judgement on the material in front of it Nobody
Went out unread No waiting step on this path Nobody, because the question was never asked

Exhibit 1. Illustrative. One message, two identity settings, and who approved each.

Read down the third column. Two settings have owners, and both owners made a defensible choice on the day: a working mailbox, and a house voice. The two decisions that carried consequence have no owner at all.

The message being accurate is what keeps this quiet. A wrong message gets corrected and the fault surfaces with it. A correct one is absorbed into the relationship, and the supplier now holds a commitment they believe the chief executive made. If she later takes a different position, the person who looks inconsistent is her.

Note what a waiting step would and would not have caught. It would have stopped the send. It would not have changed the signature, because the name at the bottom is prose in a draft, and no guard inspects prose. Identity has to be set correctly rather than filtered later.

The defenceA waiting step in front of anything that leaves under a person’s name, and one named scope that may approve it.

What happens

The run pauses, records that a decision is needed, and stops until the person whose name is on the message releases it.

What your company sees

The action was refused and the reason recorded.

Illustrative. One setting, switched. Its value is on the Register.

What it means for you

Identity is not a tone setting, and it is the one thing an agent should never be trusted to infer. Authority travels with a name whether or not the sender intended to lend it.

Three questions go to whoever built your agent, and there are three because the answers sit in three different places. Whose account does it send from. Whose name does it sign. And which of those did a person approve before the first message went out. A good answer names a mailbox, names a signature, and names the human who authorised each. If the reply is that it uses the company’s tone of voice, you asked about identity and were answered about style, and the follow-up is: show me the setting.

The safe arrangement is unglamorous. The agent sends as itself, from an address that says so, and a named person’s signature appears only on something that person read.

Next: “I will loop Jordan in”, said to Jordan.

Asked plainly

Can an AI agent send email in the chief executive's name?

It can, and it will if two settings say so. One is the mailbox the message leaves from, which is configuration in the sending system. The other is the name written at the bottom, which usually sits in the drafting instructions as a line about whose voice to write in. Neither is a judgement the model makes, and neither is checked against the other unless somebody built a check.

Who should be the sender when an AI agent writes on a company's behalf?

Whichever named human is prepared to answer for the contents, and that person should have approved the message before it left. A signature carries legal and social weight: a supplier who receives a commitment signed by a chief executive treats it as one. If nobody read the message, the honest sender is the system, under a named service address, and not a person.

What is the difference between authorisation and identity in an AI agent?

Authorisation is what the agent may do. Identity is whose name it does it under. They come apart easily: an agent can be allowed only to draft, yet still draft in a chief executive's voice from a chief executive's mailbox, so the moment a send path opens the draft goes out wearing an authority it was never granted.

Talk it through before you decide

A discovery call, no deck: your situation, the parts of the room it touches, and what you would need to decide first.