The item is on Thursday’s agenda, twenty minutes between the audit committee’s report and a property lease, and you will present it. The directors have had the papers since the weekend. You know from every other item what that means: one of them has read everything, two have read the first page, and the chair has read your covering note and formed a view of the whole proposal from its opening paragraph.
Your job in that room is narrower than explaining how an agent works. It is one decision, taken and minuted, and no director able to say afterwards that they were asked to approve something they could not describe.
The board reading path is the pack you forward before the meeting. This page is for the evening before it, and for the person who has to stand up and speak.
Four questions, asked in the same order every time
Boards are predictable about new spending of this kind, and the predictability is useful to you. Whatever the paper says, the questions arrive in one order: what is this, why are we not building it ourselves, what are we signing, and what will it cost us next year. A fifth question, about who is accountable when it gets something wrong, travels with all four, which is why it has its own section further down.
Each of the four has a page written for it, and each of those pages was written to be read cold, by somebody with no vocabulary and limited patience. That is what makes them fit for a board. Each one answers its own question completely and then stops, so a director who reads only the page touching their committee still gets a whole answer.
The order matters more than it looks. A director who has not settled what the thing is will spend the cost discussion asking what it is, in different words. A director who has not heard the case against building it will reopen that case in the middle of the contract terms. Present the four out of order and you answer each question twice. Present them in the order they are asked and each answer becomes the ground the next one stands on.
The pack carries no architecture, and that is deliberate. The design behind these pages is the Cabinet, ArkOne’s reference design for an executive agent, with eleven named parts and every setting stated once on the Register. A boardroom needs none of the eleven names. Each part sits underneath a sentence your directors already understand: it stops after a set number of rounds, it may only write to people on a list, certain steps wait for a person, and everything it does is written down where nobody can edit it afterwards.
What each page lets you say
Here are the four, with the sentence each one lets you say aloud and the question that sentence draws from the director who read the page properly. Prepare for the last column. It holds the questions you will actually be asked.
| The board asks | The page that answers it | What you can say in the room | What the director who read it asks next |
|---|---|---|---|
| What is this, actually | How a loop, a toolbox and a filing cabinet become an executive team | It is software that goes around a loop a limited number of times, using tools we allow, reading files we give it | Who decides which tools it is allowed |
| Why not build it ourselves | The Cabinet and building it in-house | The part that reasons is the small part; the guards, the ledger and the testing are most of the work | Has our own technology team costed the whole job, or the interesting part of it |
| What are we signing | Before you sign | The terms that matter say what it may send without a person, what waits for one, and what is written down | Where in the contract does each of those three sit |
| What will it cost next year | The budget conversation | The bill is a few multiplications, and we know which of them has a ceiling | Which one has none, and who is watching it |
Exhibit 1. Illustrative. The four questions a board asks, the page that answers each, and the follow-up to prepare for.
Spend tonight on the right-hand column rather than on the pages. You have read the pages. What you have probably not done is answer those four follow-ups in a sentence each, from your own company’s position, and those four sentences are the difference between an item that ends in a decision and one that ends in a request for another paper.
Two of the follow-ups are uncomfortable on purpose. The second asks whether your own technology team has scoped the whole job, and the honest answer is often that nobody has asked them to. Ask them before Thursday. If your chief technology officer has scoped all of it and still wants to build, let that case sit beside yours in the papers, because a board that has heard both chooses with more confidence than one that has heard either. The journey written for that conversation is worth reading first.
The fourth asks who is watching the multiplier with no ceiling. If nobody is yet, say so and name who will be. “Nobody yet; from the first month it is the finance director” is an answer a board accepts. A pause followed by “the vendor” is not.
The first follow-up has the plainest answer, so give it plainly. Which tools it is allowed is a list, written before anything runs, and the reference design keeps a parallel list of the people it may address: a message from somebody who is not on it is dropped unanswered. Say that the list exists, and say who keeps it.
The question that travels with all four
Whichever page is on the table, one director will ask who is accountable when it gets something wrong. It is the right question, and it deserves an answer with a mechanism in it rather than a policy.
The mechanism is one setting. In the reference design each specialist carries one word saying how much it may do on its own: act, propose or escalate. Act means it does the thing and records it. Propose means it prepares the thing and waits for a named person. Escalate means it passes the matter up without preparing anything. Every specialist starts at propose, a person has to write act before anything happens without a human, and that change goes into a record that is append only, so the entry saying who allowed it stays where it was written. The nineteenth paper sets out the whole of it, and it is the page to send the director who asks the question a second time.
That setting is also what lets a board govern something it cannot inspect. Your directors will never read the system. They can read a list of which kinds of work may act, which must propose, and who approved each change, and they can ask to see that list every quarter. A board that asks for it is governing. A board that asks for assurance instead is relying on whoever writes the assurance.
The decision to ask for
A first board item on this subject often ends in a request for more information, and the motion is a likelier cause than the room. “Approve an AI programme” asks directors to judge something with no edges, and a careful board declines to. Ask instead for a decision with edges: one piece of work, one owner, one ceiling, and a short list of what the board keeps for itself.
Suppose the proposal is a single piece of recurring work, the weekly chasing of overdue renewals, run for one quarter. A motion with edges might read like this.
| The board approves | The board keeps for itself |
|---|---|
| One named piece of recurring work, for one quarter | Adding a second piece of work |
| A named executive who owns it and reports on it at each meeting | Changing who owns it |
| A monthly spending ceiling, with the quantities behind it written beside it | Raising the ceiling |
| Every specialist set to propose and wait, with the approving people named | Moving any kind of work to act, where it proceeds without a person |
| A record of every action, kept where the auditor can read it | Anything sent in the company’s name to customers or regulators without a person signing |
Exhibit 2. Illustrative. A first motion with edges: what the board approves, and what it keeps for itself.
Every line on the left can be checked at the next meeting. Every line on the right is a decision the board has kept rather than delegated without noticing. The director who is uneasy about the whole idea can vote for it, because the part that worries them is reserved, and the keen one gets work that starts the following week.
The right-hand column has a second use. It is the governance document your audit committee will eventually ask for, already written, and dated from the day the board approved it rather than from the day somebody went looking.
Before Thursday
Two things, and neither needs you to know more about agents than you do tonight.
Send the four pages with a covering note of three sentences: what is proposed, what the board is asked to decide, and which page answers which question. Most of the board reads only the note, so the note carries the decision and leaves the explaining to the pages.
Then write your answers to the four follow-ups, one sentence each, and send them to your chief financial officer and your chief technology officer tonight. If either of them disagrees with a sentence, you want to hear it on Wednesday and not in the meeting.
If your board has less time than twenty minutes, the board reading path names the two pages to send alone. If the item becomes a standing one, the first ninety days is what your executive will need the week after the motion passes.
component: null
Asked plainly
How do I present an AI agent proposal to my board?
Answer the four questions a board asks about new spending of this kind, in the order it asks them: what the thing is, why the company is not building it, what it is signing, and what it will cost next year. Send one short page for each ahead of the meeting, carry the decision in a three-sentence covering note, and ask for a motion with edges rather than approval of a programme.
What should a board approve for a first AI agent pilot?
One named piece of recurring work for one quarter, one executive who owns it, a monthly spending ceiling with the quantities behind it, and every part of the system set to prepare work and wait for a named person. The board keeps for itself any widening of what the system may do alone, any change of owner or ceiling, and anything sent in the company's name without a person signing.
Who is accountable when an AI agent gets something wrong?
The person who allowed it to act alone. In a well-built design each kind of work carries one setting, act, propose or escalate, and it starts at propose, so a person has to change it before anything happens without a human. That change is written into a record nobody can edit afterwards, which gives a board a name and a date for every decision to let the system act unsupervised.