This paper is written for the people who have to answer for the AI bill. The CFO who cannot tie the spend to a return. The board member who sees the number rising and cannot read it. The platform or finance leader who has to make the spend legible before the next budget cycle. It answers one question. As AI moves from a feature you buy to a cost that scales with everything you do, how do you make that cost legible, attributable, and controllable, before agents make it impossible.
The Metered Enterprise, the Agentic Cost Multiplier, Intelligence Unit Economics, and the Attribution Rate are original constructs of this work, published under the license on the back cover. The framework rests on measured research into inference-cost reduction through routing [1] and budget-tiered evaluation [2]. Spend figures are framed as directional, without invented statistics.
The first AI invoice was small enough to approve without reading. The tenth was not. Somewhere between them the spend stopped being a line item and became a variable, one that rises with every new use, splits across a dozen vendors, and answers to no single owner. The CFO who asks what the company is getting for it rarely gets a clean answer, because nobody built the instrument that would give one.
This is the quiet condition most companies are now in. They run on intelligence they rent by the token, priced in a unit that has almost nothing to do with the value it produces, and the bill is about to get harder to read, not easier. The reason is agents. A person asking a model a question is roughly one call. An agent handed a task loops, pulls in context, and sets off a cascade of downstream calls, so one request becomes many, and the cost scales with how the agent behaves rather than with how many people you employ.
I wrote this for the people who have to make that spend answerable. The fix is not a freeze, which only stops the value too. It is a discipline, treating AI cost the way a good operator treats any unit economics, measured per outcome, attributed to a use, and governed with real levers. The frameworks are original. The numbers about the market are kept honest and directional. The aim is to hand you the instrument before the agents arrive in force.
AI spend has outrun the instruments built to read it. The number rises, the invoice fragments across vendors, and almost none of it is tied to an outcome. The answer is not to freeze the spend, which freezes the value too. It is to make the spend legible, attributable, and governable, as a discipline. Five findings follow.
The invoice became unreadable. You are billed per token, by many vendors, for work no one has tied to a result. That condition is the Metered Enterprise, and its defining feature is that the bill arrives before anyone can say what it bought.
Agents break the forecast. One request used to mean roughly one call. An agent loops, pulls context, and triggers downstream calls, so one request becomes a cascade. The Agentic Cost Multiplier makes spend scale with behavior, not headcount, which is why last year's forecast no longer holds.
Cost per token is the wrong unit. The question that matters is not what a token costs. It is what an outcome costs. Intelligence Unit Economics reframes the whole problem as a unit-economics discipline, measured per resolved task, not per million tokens.
The metric is the Attribution Rate. The share of AI spend you can trace to a specific use or outcome. Most organizations cannot state theirs, and what you cannot attribute you cannot govern, defend, or cut without also cutting the value.
Legibility is a build, and it funnels to ownership. You make spend legible in phases, Meter, Attribute, Govern. The deepest lever is taking the routine majority of work off the meter and onto infrastructure you control, where cost becomes fixed and predictable.
Ask most companies what their AI actually costs and you get a number. Ask what that number bought and the room goes quiet. The spend is real and rising, but it arrives as a stack of invoices priced per token, spread across model vendors, tooling subscriptions, and infrastructure, and almost none of it is tied to a use anyone can name. The bill is legible as an amount and illegible as a decision.
That condition has a name. The Metered Enterprise runs on intelligence it rents by the token, in a unit that has little to do with the value produced, with no single owner and no line from a dollar to an outcome. It is not that the spend is too high. It is that it cannot be read, and what cannot be read cannot be governed, defended to a board, or cut without cutting the value along with it.
The AI bill is legible as a number and illegible as a decision, and the gap between the two is widening.
Last year's budget assumed a simple shape. A person asks a model a question, the model answers, you are billed for roughly one call. You could forecast that, because it scaled with how many people used it. That assumption is expiring, and agents are what expires it.
Hand an agent a task and it does not make one call. It plans, retrieves context, calls a tool, reads the result, reconsiders, and calls again, often several models deep, sometimes spawning other agents to do pieces of the work. One request becomes a cascade. This is the Agentic Cost Multiplier, and its consequence is that spend now scales with how the agents behave, not with headcount, which is exactly the variable your forecast did not have.
You used to pay per person who asked. Now you pay per step an agent decides to take, and it decides a lot of them.
The reason the bill is unreadable is that it is denominated in the wrong unit. Vendors sell tokens because tokens are what they can meter. But a token is an input, not a result, and a CFO cannot run a business on the price of an input that may or may not have produced anything. Managing cost per token is like managing a sales team by the word count of their emails.
Intelligence Unit Economics prices the work by the unit that matters, the completed outcome. Tokens roll up into calls, calls into finished tasks, tasks into booked outcomes. Only the last is a unit a CFO can govern, because only it ties a dollar to something the company kept. The reframe is moving your attention down that ladder.
Vendors meter the input. You manage the outcome, the only unit that governs.
| Unit | What it measures | What it hides |
|---|---|---|
| Tokens | Raw model throughput | Nothing about whether the work was any good |
| Calls | How busy the system was | Whether the calls were necessary or duplicated |
| Completed tasks | Work that actually finished | What the finished work was worth |
| Cost per outcome | Spend against a result you booked | It hides nothing. This is the unit to manage |
A discipline needs a number. For AI spend the number is the Attribution Rate, the share of your AI spend you can trace to a specific use or outcome. It is the single figure that tells a CFO whether the spend is managed or merely incurred, and most organizations cannot state theirs, which is the whole problem in one statistic.
The Attribution Rate works like a credit score for AI spend. A low rate does not mean you are spending too much. It means you cannot tell which of the spend is working, so you cannot cut the waste without also cutting the value. Raising the rate is the point of the entire exercise, because every dollar you can attribute is a dollar you can defend, forecast, or redirect on purpose rather than by guess. Measuring it well, under real budget limits, is itself a skill [2].
Once you can attribute spend, you have levers, and they are not exotic. Route the easy majority of work to cheaper models and keep the frontier for the hard minority [1]. Cache what repeats. Stop over-provisioning context and model size for tasks that do not need them. And for the routine bulk, take it off the meter entirely by running it on infrastructure you own, on a local runtime such as Zorp [3], where the cost stops being a variable and becomes a fixed line you control.
No single lever is the answer, and the biggest one is the one most programs reach for last. Owning the routine majority turns the largest slice of a variable, metered bill into a fixed cost you can plan around. The levers stack, and stacked they change the shape of the whole number.
Route the easy work, cache the repeats, right-size the rest, and own the majority.
You do not need a platform rebuild to make spend legible. You need three moves in order. Meter. Instrument the spend first, by use and by team, so the invoice stops being a single opaque number. You cannot manage what you cannot see, and most organizations have never actually looked.
Attribute. Tie that metered spend to the uses and outcomes it serves, and compute your Attribution Rate. Govern. With spend visible and attributed, apply the levers, set budgets by use, route by difficulty, and move the routine majority onto owned infrastructure. Each phase has a gate. Do not pass it until the number is real.
You do not fix everything at once. You find the first thing that is broken and fix that. The order is fixed, because each move depends on the one before it. You cannot attribute spend you cannot see, and you cannot govern spend you cannot attribute. Find the first No in the chain, and that is where you start.
Make the spend legible before agents make it unforecastable. See it, tie it to outcomes, control it with levers, and take the routine majority off the meter onto infrastructure you own.
| Move | Question | Yes | No |
|---|---|---|---|
| 1 METER | Can you break your AI spend out by use and by team, not as one number? | ||
| 2 METER | Do you know your single biggest driver of AI cost this month? | ||
| 3 AGENTIC | Have you measured what an agent run costs versus a single call? | ||
| 4 UNIT | Do you track cost per completed outcome, not just cost per token? | ||
| 5 ATTRIBUTE | Can you state your Attribution Rate, the share of spend tied to an outcome? | ||
| 6 ATTRIBUTE | Is that rate rising quarter over quarter? | ||
| 7 LEVER | Do you route easy work to cheaper models and keep the frontier for the hard minority? | ||
| 8 LEVER | Do you cache repeats and right-size context and model choice? | ||
| 9 OWN | Do you run any routine workload on infrastructure you own, off the meter? | ||
| 10 GOVERN | Does a budget exist per use, with an owner, that someone reviews? |
A CFO can hand these ten questions to the team that owns AI spend. An answer with a number behind it passes. An answer that is a shrug does not.
This paper is a conceptual and practitioner contribution. The Metered Enterprise, the Agentic Cost Multiplier, Intelligence Unit Economics, and the Attribution Rate are original constructs of this work. They rest on measured research where it bears on the levers, the economics of routing queries across model tiers to cut cost [1] and the evaluation of models under real budget limits [2]. The ownership lever points to a local runtime for running the routine majority on hardware you control [3].
The figures about AI spend are framed as directional, on purpose. The aim is not to publish a market-sizing number or an invented percentage. It is to give the reader the instrument, a unit, a metric, and a set of levers, so they can measure the real figures in their own house, which are the only ones that matter.
Arjun Jaggi, The Metered Enterprise, Enterprise AI Research, White Paper No. 11, arjunjaggi.com, November 2026.
Arjun Jaggi writes and advises at the intersection of enterprise AI strategy, the economics of adoption, and the operating work of turning AI into results. His frameworks turn the hardest questions of running AI at scale into terms a board can decide on and a team can execute. Engage at arjunjaggi.com.