Your head of technology has the licence figure on a slide, and it is not small. Sixty seats across finance, sales and operations, renewed annually, and the report attached is encouraging. Drafting is faster. The spreadsheets your analysts build come together in a morning rather than two days. When you ask what else the company needs, the reasonable answer arrives: you have already bought this, extend it. What is being offered is a tool that makes a present person better at their job. What you asked about was the work that goes unstarted, and no number of seats reaches a Friday afternoon when the office is empty.
The seat is attached to a person
The seat earns the praise on that slide. It is good at making one person faster inside one application, and it holds a property no agent can match: the human stays in every loop. Nothing leaves without somebody pressing send, nothing is decided without somebody reading it first, so the question of who authorised what never has to be answered. Your general counsel has no interest in a copilot seat, and that is a feature you are paying for whether you noticed it or not.
The property that limits it is the same one. A seat is attached to a person and an application, so its output follows attendance. The analyst who builds a model in a morning rather than two days has to be there for that morning.
An agent is attached to work rather than to a person. In the Cabinet, ArkOne’s reference design for an executive agent, the fixture that makes this possible is the Clock, a scheduler that wakes every thirty seconds and claims a due job in one indivisible step, so the job fires once and only once whether or not anybody is logged in. What survives between those wakings is the Minutes: decisions written after a meeting has ended, then read back at the start of the next one, the eight most recent alongside every initiative still open. A seat has no equivalent, because it has never needed one. When the person comes back tomorrow, they are the continuity.
The six rows
Suppose a company invented for this exercise: two hundred people, sixty seats, a commercial team of nine holding about forty live accounts. What the seats produced is real. Proposals go out faster and read better. What nobody has costed is that follow-up still slips in the weeks the team travels, and August is quiet in a way nobody chose.
| The row | A copilot seat | The Cabinet |
|---|---|---|
| Who decides | The person in the chair, on every action; the seat suggests and waits | A specialist at its authority setting, shipped at propose and wait; only the chief executive’s scope approves anything at all |
| What it can touch | Whatever its holder can touch, inside the application it lives in, at the moment they act | Named systems through named tools, on its own schedule, behind outbound rules that run in code |
| What it remembers | The document, the thread, the session; the person carries the rest | Minutes written after the work and read back at the next start, plus a morning brief that synthesises rather than re-lists |
| What it costs to run | A fixed sum per seat per year, whatever the usage; predictable and often under-used | Follows the work, including work nobody asked for. Predictable only once the loop has a ceiling |
| What happens when it is wrong | The holder sees it, fixes it, and nobody outside the room learns of it | Something may have left already. What you hold is the ledger: append only, corrections added as entries, originals standing |
| Who owns it afterwards | The licence administrator; the habit belongs to each holder | Instructions, authority settings, the roster it may address, and the ledger, each with a named owner or none |
Exhibit 1. Illustrative. The six rows that separate a tool sitting beside a person from one that carries work between them.
The first row is where a board conversation turns. A seat’s cost is headcount arithmetic you can put on a slide, and its risk is close to nil because a person authorises every action. That combination is comfortable, and it is why sixty seats were approved without a governance discussion.
The fourth row is the one that costs you quietly. A seat bills the same in the month a holder opens it two hundred times and the month they open it twice, so a fleet of seats has a floor and no ceiling on value left unclaimed. An agent inverts that: the bill rises with work done and falls to nothing when there is none, which is comfortable in a different way and uncomfortable in the way that counts, because the work it does is work nobody was watching.
When the seat is the better buy
There is a common shape of company where seats are plainly the right purchase and an agent would be a mistake.
The work is skilled, the judgement is the point, and the bottleneck is production. Your analysts know what the model should say and typing it is the slow part. Your salespeople know the client and assembling the proposal is the slow part. In that company, every hour of value is created by a person exercising judgement, and a tool that removes typing from around that judgement is close to pure gain. Buying authority instead would mean building a roster, a ledger and a set of outbound rules to govern actions you would rather a human took anyway.
The question that separates the two situations is about attendance rather than capability: which of the things that went undone last quarter went undone because nobody was there?
Go through the list honestly. The proposal that was late because it took two days to build is a production failure, and a seat fixes it. The renewal nobody raised because the person who tracks renewals was on leave for three weeks is an attendance failure, and no seat reaches it, because a seat is that person’s tool and that person was in Portugal. If your list is mostly production failures, renew the seats and stop reading. If it is mostly attendance failures, you have been buying multipliers for a problem that is about coverage.
Take last quarter’s list of things that slipped and mark each one production or attendance. The mix is your answer, and it is a more honest business case than any licence utilisation report. If attendance dominates, the six rows above are what you take into the vendor conversation, and the first row is the one to press: ask who decides, and ask to see where that setting lives. The Cabinet and a chatbot runs the same six rows against a tool with no authority at all.
component: null
Asked plainly
Is a copilot licence the same as an AI agent?
No. A copilot is attached to a person and an application, and it produces value while that person is working inside that application. An agent is attached to a piece of work. It has a schedule of its own, an authority setting saying what it may do without asking, and rules governing what it may send outward. The practical difference shows up in the hours nobody is at a desk, which is where a copilot produces nothing at all.
We already pay for copilot seats. Why would we buy anything else?
Because the two purchases fix different problems, and paying for one does not reduce the case for the other. Seats make the work your staff already do faster and better, and that is worth paying for on its own. What seats cannot reach is work that nobody currently starts: the renewal that passed unnoticed, the report nobody assembled, the client who was not chased on the Friday everyone was travelling. If your unfinished work is mostly of that kind, more seats will not touch it.
How do I tell whether copilot seats are earning their money?
Look at usage per seat rather than at the total licence count, and look at what the heaviest users are doing with it. A seat pays for itself when one person opens it repeatedly for a task they used to do slowly, so a small number of intense users is a healthy pattern and broad shallow usage is not. The finding worth acting on is a seat bought for a role that opens it twice a month, because that is a licence funding a habit nobody formed.