The board pack closes on Friday and your line for this is still a single figure with a question mark beside it. Two vendors have given you numbers. One is a price per person per month, tidy, the sort of line an auditor never queries. The other is a range, with the word usage in front of it, and a note saying it depends on how much you use the system.
Neither is defensible in the room you are walking into. Your chief financial officer will not ask whether the figure is high. She will ask what makes it move, and what it looks like if the answer to that is twice as much of it. You have been in this position with a logistics contract and a software renewal, and both times the useful conversation started when somebody put the quantities beside the rate.
That is the conversation available to you here, and it has not happened yet, because nobody has shown you the quantities.
A bill with four multipliers in it
Set aside the unit price for a moment. Almost nobody in this market is dishonest about the rate, and the rate is not what will surprise you.
A running bill for an agent comes out of a small number of multiplications. How many rounds a piece of work takes before it finishes. How many specialists get consulted in each of those rounds. How much material is read on each of those calls. How many lookups reach outside your company. Multiply those together and you have the shape of the invoice, and every one of them is a setting somebody chose before you were shown a screen.
Which means your job in this meeting is narrow and doable. You are not there to negotiate the rate down. You are there to find out which of those multipliers has a ceiling written into the design and which one is open-ended, because an open-ended multiplier is what turns a comfortable month into the invoice that gets escalated to you.
The design behind these pages is the Cabinet, ArkOne’s reference design for an executive agent, and every one of its settings is written down on the Register, the standing page of the whole design. One of them bounds the first multiplier: the loop may go around at most fifteen times before the work must end. That is worth having, and here is the limit in the same breath. A cap on rounds bounds how long one piece of work can run. It does not bound what those rounds cost. A round that consults one specialist and a round that consults every specialist both count as one round, and the second costs many times the first. If you take one sentence from this page into the meeting, take that one, because it is the sentence a well-intentioned vendor will get wrong in front of you.
The second multiplier is the room itself. In the reference design a piece of work may consult every specialist on the roster at once, ten of them, and the ceiling is the roster rather than some smaller number. That is a deliberate choice with a cost attached, and what actually restrains it is the same thing that restrains a meeting in your company: every opinion is a reply somebody pays for and reads.
The third is the one nobody mentions and the one that moves the bill most. What the agent reads before it answers.
The four multipliers, and what a vendor says when one is loose
Here are the four, each with what bounds it in the reference design and how the answer sounds when nothing does. The fourth column is the sentence you will actually hear, and none of the four is dishonest.
| The multiplier | What it is | Bounded in the reference design by | Loose sounds like |
|---|---|---|---|
| Rounds per piece of work | How many times it goes around before it must answer | A ceiling of fifteen rounds, then the work ends | It keeps going until it has a good answer |
| Specialists per round | How many advisers a single round consults | The roster, and nothing smaller; the restraint is that each reply is paid for and read | It asks whoever is relevant |
| Material read per call | The standing brief, the tool list, the documents, the conversation so far | Two standing blocks that hold their shape, a sorted tool list, and a relevance line beyond which a document is discarded rather than read | It has access to all your documents |
| Lookups outside the company | Searches of the open web during one piece of work | Two per piece of work, each billed at the provider’s published rate | It searches when it needs to |
Exhibit 2. Illustrative. The four multipliers behind a running bill, what bounds each one, and how a vendor answers when nothing does.
The third row is where the money is, and it is the row that gets one line in a proposal. Everything the agent knows at the moment it answers has to be handed to it on that call, because the model keeps nothing between one call and the next. The standing description of who it is and where it is working, the list of what it may do, the documents it pulled, every earlier exchange: all of it, every time, for the life of the piece of work.
So the same material read repeatedly is the largest single line on most of these invoices, and one property of the design matters more to your budget than anything else here. Material identical to what the previous call carried can be read from a copy the provider kept, at a much lower rate than reading it afresh. According to Anthropic’s published pricing, that cached rate is roughly a tenth of the fresh one, and the Register records the same ratio. Only a handful of places in a request can be marked as the boundary of what may be re-read from that copy, four of them in the reference design, so the choice of where they go is a budget in itself: two go on the blocks of the standing brief, one on the list of available tools, and one is kept back for the conversation. The tool list is sorted by name before every call, so its order never drifts.
That sorting sounds like housekeeping. It is the difference between the low rate and the full one, on everything below it, on every call. A list assembled in whatever order things happened to load looks identical on screen and never matches the copy, so nothing after it matches either, and the saving is gone without anything appearing to break. The eleventh paper walks the arithmetic of one moving line at the top of the page.
The fourth multiplier carries its own limit. Outside lookups are capped at two per piece of work, which is a real ceiling and a real constraint on what the system can find out, because an agent free to keep searching would sometimes answer better. Two is a judgement about where the cheaper mistake lies, and it is yours to move once you know the price of moving it.
The same month, twice
Suppose a company of about three hundred and twenty people, invented for this exercise, running an agent across renewals and client correspondence for one month. Same volume of work in both columns, same specialists, same documents. The figures below are illustrative and modelled from the cached-read ratio, not measured from anything running.
| For one month, illustrative | The tool list holds its order | The tool list drifts |
|---|---|---|
| Pieces of work completed | The same in both columns | The same in both columns |
| What the agent read, billed | Mostly at the low cached rate | Almost entirely at the full rate |
| What the agent wrote back, billed | Unchanged; replies are never cached | Unchanged; replies are never cached |
| The modelled monthly bill | Call it one unit | Call it several times that unit |
Exhibit 3.
Nothing in the right-hand column is broken. The work completed, the quality of the replies and the screens everybody looks at are identical, which is exactly why nobody finds this in a demonstration and somebody finds it on an invoice in the fourth month.
And the gap only opens on long work. A conversation of two exchanges has almost nothing to cache, so a short demonstration shows no difference at all. Your own multiple would depend on how much standing material the system carries and how long a typical piece of work runs, and the ratio belongs to the provider and can change. The direction does not change, and no other setting recovers what a drifting list spends.
The third row is the one to read carefully, because it bounds the saving. Caching reduces what is paid for reading and does nothing for what is paid for writing, and every reply, including every answer a specialist gives inside a round, is billed at its own rate. So the second multiplier sits outside the saving entirely. A design with excellent caching and an unbounded habit of consulting everybody has one half of the bill under control.
There is a matching restraint on the depth of thinking. The reference design holds its four deep-reasoning specialists at the low effort setting, on the modelled basis that higher settings cost roughly three times the time without a gain anybody has been able to measure, and the setting moves the moment a standing examination can show what the higher one buys. The eighth paper puts the same question at two settings side by side.
Three questions to put in writing
Send these as an email before the figure goes anywhere near your board, because you want the answers in a document rather than in somebody’s recollection of a call. All three are procurement questions.
What are the quantities behind the figure you have given me? The good answer is four numbers: rounds per piece of work, specialists consulted per round, how much material is read on a typical call, and outside lookups allowed. A vendor who has measured their own system produces those in a day, from logs, with a typical month and a heavy one beside each other. The answer that arrives instead is that it depends on your usage patterns, which is true and is the result of the arithmetic rather than the arithmetic. Your follow-up is to make it specific: take one piece of work your company would actually do, one renewal reply, and give me those four numbers for that.
Which of those quantities has a ceiling I can read, and which is open-ended? The good answer separates them without being pushed, and says where each ceiling is set and who may change it. The answer to press on is that everything is configurable. Take it at face value, because it is generally true, and then ask the only question that matters about it: what is it set to today, in the instance you would be running for me, and does changing it need your engineers or mine. A ceiling nobody can name the current value of is not bounding anything.
Show me the bill for one piece of work with caching working and with it broken. The good answer is two figures from a log with the share read from the copy beside each. A vendor who can produce that has measured the thing this page is about. Two replies should send you back to your own numbers. One is that caching is the provider’s business and not theirs, which is half true: the provider caches what matches, and the vendor’s arrangement of the material decides whether anything matches. The other is that a flat fee per person makes their running costs none of your concern. That is a genuine service and it moves who carries the risk without changing how the material is assembled, so the follow-up writes itself: what happens to that fee at renewal, and what happens to it if my volume doubles.
Notice what none of those asks about. Not the model, not the rate per unit, not the roadmap. Three questions about quantities, and a vendor who answers all three has told you more about how carefully the thing was built than any demonstration could.
Ask for one meeting’s invoice
Before Friday, do one thing that needs no understanding of any of this. Pick a single piece of work the system did last week, a specific one with a name on it, and ask the vendor for the cost of that one piece of work broken into what was read and what was written, with the rounds counted. Ask for it by the end of the week.
The reply is the finding, whatever it contains. A vendor who sends a short table has instrumented their own system and can help you budget. A vendor who sends the monthly total again, or explains that costs are not tracked at that level, has told you that nobody on their side knows which of the four multipliers is loose, which means the first person to find out will be your finance director, in the fourth month, from an invoice.
Which is the whole value of asking for one named piece of work rather than a better rate. A rate you can argue about tells you nothing about volume, and the multiplier nobody has measured is the one that decides what you pay when your volume moves. You are buying the arithmetic, not the price, and a vendor who can show you the arithmetic on one invoice can show you what next year looks like. When the quantities do come back and a contract follows, the terms that decide how the arrangement actually goes are not on the commercial page either, and the four of them worth your counsel’s time are worth an evening.
component: null
Asked plainly
How do I budget for an AI agent when the vendor quotes usage-based pricing?
Ask for the quantities behind the figure rather than a better figure. A running bill comes from how many rounds a piece of work takes, how many specialists each round consults, how much material is read on each call, and how many outside lookups happen. Each of those is a setting somebody chose, and each either has a stated ceiling or does not. A quote with no quantities attached is the result of arithmetic you were not shown, and it will move when your usage does.
Is per-seat pricing for an AI agent safer than usage-based pricing?
It is more predictable for you and it hides more. A seat price moves the running cost onto the vendor, which is a real service and a reason the figure is often high. What it does not do is tell you whether the work underneath is assembled carefully, and an arrangement whose cost is invisible to you is one you cannot argue about at renewal. Ask for the quantities anyway, and ask what happens to the seat price when your volume doubles.
Does capping how many times an AI agent loops control the cost?
It bounds one of the multipliers and not the bill. A ceiling on rounds means a single piece of work must end rather than run all night, which is worth having. Inside each of those rounds the system may still consult several specialists, read a long document, and make outside lookups, and a round that consults everybody costs many times a round that consults one. A round cap is a limit on duration, and a cost limit is a separate thing to ask for.