The renewal notice came through your finance inbox rather than your own, which is how you found out it was due. Per-person assistant licences, bought a year and a half ago for part of the company or all of it, at a figure that was easy to approve at the time and is now a line somebody queries. Attached is a usage report you have read twice without learning anything from it.
What you know without the report is this. Two or three people in the company talk about it unprompted, and one of them has stopped doing a thing you used to hear complaints about. Everybody else says it is useful when asked directly and mentions it at no other time. Nobody has produced a number, and you suspect that the number, if it existed, would not settle the question either.
Somewhere in the same fortnight a supplier has told you that what you actually need is an agent. You have no idea whether that is a different thing or the same thing with a better margin on it.
Two purchases, and the question that separates them
The licences were not a mistake, and a page that opened by telling you they were would be wrong about the product as well as rude about your judgement. What you bought sits beside one person and makes that person faster at their own work. It is there when they are working and gone when they are not. It knows what they show it. It remembers what that one person’s session held, and it asks before it does anything, because it does nothing unasked. A tool with those four properties is worth paying for, and the reason your general counsel has never once asked you about it is that every one of those properties is a safeguard you bought without noticing.
Each of those properties also sets a boundary, and the boundaries are structural rather than a matter of maturity or of the vendor trying harder next year.
It acts when a person is there, because a licence has no clock of its own. Nothing happens at three in the morning, not because the vendor forbade it but because there is nothing to start it.
It knows what its user knows. There is no roster of your company’s people inside it and no shelves of your company’s material, so it cannot consult what nobody showed it in that session.
It remembers a conversation rather than a company. What your sales director settled with it on Tuesday does not reach your operations lead’s Thursday, because the two sessions are separate rooms.
And authority is the user’s own. There is no setting for what it may do unasked, for the same reason there is no brake on a parked car.
Against that, the design behind these pages is the Cabinet, ArkOne’s reference design for an executive agent, and it carries a named fixture for each of those four boundaries. A scheduler that wakes on a fixed interval, every thirty seconds in the reference design, and claims a due job whether or not anybody is logged in. A roster of the people it may address at all, and four shelves of material kept apart: house knowledge, the client’s documents, recorded failures, recent research. Decisions written down once the work has ended and read back at the start of the next piece, the eight most recent alongside every initiative still open. And one authority setting against each of the ten specialists it consults, which is act on its own, propose and wait, or escalate to a person, shipped at propose and wait until somebody writes otherwise. Every value sits on the Register.
So the renewal question is not seat or agent. It is which of the work you are paying for needs a person present, and which needs to happen whether anyone is present or not. Those are two purchases and both can be correct. Set against each other row by row, the two run to six fixed rows; what follows here is the renewal rather than the comparison.
The work, split by whether anyone has to be there
Four kinds of work, in the order they appear in a week. What decides the column is attendance rather than difficulty, which is why the hardest row is served by the thing you already own.
| The work | Does a person have to be there | Which purchase serves it | Why |
|---|---|---|---|
| Producing something a person is already making | Yes, that is the whole of it | The licence, clearly | The judgement is theirs and the typing is the slow part; nothing else is needed |
| Reading and summarising so one person can decide | Yes, and the decision follows within minutes | The licence, clearly | The output is a précis for an audience of one, needed at the moment they ask |
| Carrying a decision already made outward | Not for the carrying, only for the deciding | Either, and an agent once the volume is steady | Four messages, two updates and a record entry, each easy and none finished |
| Noticing that something is due and starting it | No, and nobody currently does | An agent, or nobody | A thing nobody starts is invisible to a tool that waits to be opened |
Exhibit 2. Illustrative. The work split by whether a person has to be there, and one company's licence count against its weekly openers.
The first two rows are where the licences earn their money, and they are more of most weeks than anybody admits. An analyst who knows exactly what the model should show and loses a day to building it is not short of judgement, and a tool that removes the building is close to pure gain. Reading a long file so a decision can be made in ten minutes is the same shape: the entire output is a summary for one reader, wanted at the moment that reader asks, which is precisely when a licence is open.
The fourth row is the one your usage report cannot see, because nothing appears in a usage report that nobody did. The renewal that passed its date. The report nobody assembled because the person who assembles it was travelling. The client nobody chased on the Friday the office emptied. Each of those is a thing your company decided to do and then did not start, and no number of licences reaches it, because a licence is somebody’s tool and that somebody was elsewhere.
The third row is where the two purchases genuinely overlap, and it is the honest cost of running both. If an agent drafts the renewal replies your account team currently drafts with their own licences, then the licences in that team are buying less than they were. Size it rather than argue about it: name the teams whose work an agent would take over, count the licences held inside those teams, and ask whether each of those people still has enough of the first two rows in their week to justify a licence of their own. The ones who do not are a saving you can name on the same page as the new cost, and neither supplier will do that arithmetic for you.
The renewal arithmetic
Suppose a specialist recruitment and executive search business, invented for this exercise, with about two hundred and sixty people across three offices. It bought ninety licences eighteen months ago, one for everybody in the commercial, research and finance teams. The figures below are illustrative rather than measured from anything, and the only one that matters is the gap between the first two rows.
| For one month, illustrative | The count | What the count is telling you |
|---|---|---|
| Licences bought and billed | Ninety | The figure on the renewal notice, unchanged since the day it was signed |
| People who opened theirs in a normal week | Around thirty | The real size of the habit, and it stopped growing after the first quarter |
| Where the weekly openers sit | Almost all in research and proposal writing | Producing and summarising, which is the work a licence is built for |
| What the sixty quiet licences have in common | Roles whose week is chasing, coordinating and following up | Work that is mostly waiting, and waiting is not made faster by typing faster |
Exhibit 3.
The temptation with a table like that is to read it as sixty people who did not adopt the tool, and to spend the next quarter on training. Look at the fourth row before doing that. The quiet licences cluster in roles whose week is made of chasing a document promised on Monday, coordinating three diaries, and following up the thing that went quiet. A person in that role does not have a production problem, so a tool that removes typing from around their judgement removes almost nothing, and they will use it politely for a month and then stop.
Which turns a disappointing adoption number into a finding about the work. The thirty weekly openers are telling you the licences were correctly bought for research and proposals and should probably be renewed there without further discussion. The sixty quiet ones are telling you that the largest block of hours in that company is spent on work that needs somebody to be there and paying attention, and that buying a faster pen for it was the wrong purchase rather than the wrong people.
One caution about the second row, because it is the number that will get quoted in the room. A licence opened twice a week by somebody who uses it for the single hardest thing they do may be worth more than one opened daily for trivia, and an open count cannot tell those apart. Ask what the heavy users are doing with it before you read too much into the count, because the answer is usually one specific task and it is usually worth knowing.
What to ask, of both of them
You will have two conversations before the renewal date, one with each supplier, and the useful question is different in each. Ask the licence vendor for a count. Ask the other one for a boundary.
Ask your licence vendor how many of your people opened theirs in a normal week last month, and what the heaviest users were doing with it. A vendor who takes this seriously produces both in a day, because the first is a count they already hold and the second is the thing their own product team lives on. The reply to press on is a report of total interactions or messages sent, which rises whenever anybody is merely curious. Your follow-up makes it people rather than events: how many distinct people, in how many distinct weeks, and give me the twenty heaviest with their job titles. If the answer is that usage is not broken out per person for privacy reasons, accept it and ask for the count by team, because a team is not a person and the objection does not survive the question.
Then ask whoever is selling you the other thing what it does in the hours nobody is present, and say plainly that you already hold licences and want to know what is left over. The good answer starts with what your licences already cover, declines to sell against them, and then names the fourth row of the first table above with the mechanism that reaches it: a scheduler on a fixed interval, an authority setting per specialist, and rules that decide in code what may leave the company. The answer to distrust is the agreeable one: that an agent does everything your licences do and more, so you can consolidate. Consolidate on that advice and you will cancel the licences in research and proposals, where they are working, in order to buy something aimed at the hours nobody is at a desk. You lose a gain you already have and you pay for it twice. Your follow-up is one sentence: name the teams in my company where this would replace a licence, and name the ones where it would not.
There is a second suspicious reply from that side, and it is quieter. Where you are told that the running cost is comparable to a licence per person, ask what the figure does when the work doubles, because a bill that follows the work has no ceiling of its own and a licence has nothing but a ceiling. Ask which model answers which kind of question, and who chose. In the reference design the choice is made per call rather than once for the whole arrangement, and a capability one provider will not accept is stripped out before the request leaves so that routing elsewhere cannot turn a working call into an error. Ask also what it is allowed to look up outside your company, which in the reference design is two searches for a single piece of work, each billed at the provider’s published rate. The eighteenth paper explains why the routing decision belongs to each call rather than to the arrangement as a whole.
Say the last part of this out loud in whichever meeting settles the renewal. You are not choosing between two products. You are deciding how much of your company’s unfinished work is unfinished for want of speed and how much for want of attendance, and that split is yours to know rather than either supplier’s to tell you.
The list, and one line drawn across it
Before the renewal date, open your own licence administration screen and export the per-person list for last month. Not the summary the vendor sent. The list, one row per name, with an open count against each. It takes a few minutes and needs no permission from anybody.
Sort it by the count, highest first, and draw a line where the counts drop below roughly once a week. Then read the job titles below the line, out loud if there is nobody in the room, and notice what they have in common. If what they share is the kind of work they do rather than anything about them as people, the list has just named a category, and a category is a thing a purchase can be aimed at.
Keep both halves of the list. The half above the line is your renewal case, written by the people using it, and it is a better argument than any utilisation report. The half below is the specification for the other conversation, and it is worth taking into that conversation unedited, because a supplier who can look at those titles and say honestly which of them an agent would not help either has told you something a demonstration cannot.
What the half below the line is not is a list of people who failed to adopt something. It is a category of work, and the reason it stayed quiet is that the work in it was never the kind anybody speeds up by typing faster. Cancelling those licences saves money and changes nothing about the work, which is the point: the hours are still there, still unfinished, and still waiting for a different purchase or for nobody. When that conversation turns to what the different purchase costs to run, the budget conversation is the one to have before the figure reaches your board.
component: null
Asked plainly
We already pay for per-person AI assistant licences. Do we still need an AI agent?
It depends on which of your unfinished work needs somebody present. A per-person licence sits beside one member of staff and makes that person faster at their own work, which is a real gain and only available while they are at their desk. An agent is attached to a piece of work rather than to a person, so it has a schedule, an authority setting saying what it may do unasked, and rules governing what may leave the company. If what went undone last quarter went undone because nobody started it, more licences will not reach it.
What will a vendor's usage report not tell me before a renewal date?
It cannot show you work that nobody did, which is usually the thing the renewal should turn on. A usage report counts what happened inside the product: sessions, messages, documents drafted. The renewal that passed its date, the report nobody assembled while its author was travelling, the client nobody chased on a Friday: none of that appears, because no licence was opened and no event was generated. So a report can tell you whether the licences you bought are being used, and it cannot tell you whether they were the right purchase. For that you need to name the work your company decided to do last quarter and then did not start, and ask whether a faster tool for a present person would have reached any of it.
If we buy both, are we paying twice for the same thing?
Partly, and the overlap is worth sizing rather than arguing about. The part that genuinely overlaps is drafting and summarising for people who are present and working, and if an agent drafts the replies your staff used to draft with their own licence, the licences in that team are worth less than they were. Size it by naming the teams whose work the agent would take over, counting the licences held there, and deciding whether those licences are still worth buying for the rest of what those people do.