AI economics · The argument

An agent decides how much work to do. Your cost controls assume a person did.

Most technology spend starts with a person who provisions, buys or approves. An agent is different. A person may start the workflow, but an agent decides which steps to take, which tools to call and when to stop, and every step can cost money.

The problem

Your controls were built for a world where a person decided

Core assumptions sitting in tools enterprises use to manage technology cost today. Cost is static. Spend stays inside IT. Planning happens once a year. AI breaks all three at once.Cost scales with usage. Spend arrives through business units, cloud, embedded software and expense lines. And agents work at machine speed against a plan that was written months ago.

One request becomes a chain of spend

A chatbot has a visible link between a request and a reply. An agent does not. It breaks a goal into tasks, passes some to other agents, retrieves documents, runs code, calls enterprise systems and checks its own work. It can run exactly as designed and still cost more than the result is worth.

The multiplier is invisible in the request

Engineers now classify AI workloads the way they classify algorithms, by how consumption grows and not by what one call costs. A cached lookup is constant. One model call per request is linear. Reasoning steps and tool calls multiply it. An orchestrator that starts sub-agents multiplies everything beneath it. A retry loop with no stopping rule has no ceiling at all. The expensive patterns exist in production today, and they are found on the invoice.Big-T Notation, published by the Tokenomics Foundation, 2026.

The question is no longer who used the AI. It is who owns the economic behavior of the agent.

5 to 30 times

An agentic workflow uses five to thirty times the tokens of a chatbot query for the same task.Gartner, March 2026.

3 times

By June 2026, enterprises were at three times their full-year forecast for token spend, with half the year left.FinOps Foundation, FinOps X 2026.

98%

Of FinOps practitioners now manage AI spend.FinOps Foundation, State of FinOps 2026.

21% and 23%

CIOs are accountable for 21% of AI delivery and 23% of AI funding.Gartner, May 2026.

What the industry has tried

Four fixes exist. None says who owns the agent.

Cloud FinOps, extended

Cloud cost is attributed through tags applied when a resource is provisioned. An AI call carries no team, product or owner by default, so attribution has to be designed into the agent. It cannot be applied to the bill afterward.

Token monitoring tools

A new set of tools watches token consumption by service. Tokens are how AI output is metered, the way kilowatt hours meter electricity, and nobody describes the economics of the power industry as the discipline of reading meters. A team can cut token cost and leave the far more expensive workflow design untouched.

The annual budget and the monthly close

Both were built for licenses and headcount, costs that are committed and follow a known schedule. An agent has no schedule. Its cost turns up on the invoice, weeks after the chance to move money has closed.

Frameworks that name the job, but no system to do it

The FinOps Foundation now recommends clear ownership, granular cost tracking, thresholds and governance across functions. Gartner has named AI-specific financial management: continuous monitoring of cost, value and risk per successful outcome. Neither names a system that does it.

“Token economics is fundamentally more abstract and more opaque than anything we have managed at this scale before. It requires a different operational muscle than the one the industry built for cloud.”

Arvind Joshi, COO and CFO of Global Technology, JPMorganChase, 2026.

What it takes

Give every agent an owner, a purpose and a measurable result

Treat every agent as an economic object. It gets an identity, an owner, a purpose and a result you can measure. That is what lets cost be attributed even when several agents share one model, and it is what makes the six moves below possible. They are the six verbs of MOTEF, the MagicOrange Technology Economics Framework.

See

Every agent, model call and GPU hour sits in the same model as the cloud, software and people around it.

Price

Cost per inference and cost per successful outcome, loaded with tokens, platform, data and labor.

Plan

A switch of model, or a doubling of volume, runs as a scenario and not as a rebuild.

Optimize

Compare what each agent costs with what it produced, then scale it, redesign it or retire it.

Control

Watch how fast spend is climbing, not only the total. Set budgets and thresholds before the invoice arrives.

Prove

Set cost per outcome against the value of the outcome, and book the value to a ledger line.

Agent cost sits across the whole estate

An agent uses cloud services, software, model APIs, data platforms and shared infrastructure. All of that lands in one governed model across seven spend categories. The question changes from which agent is expensive to whether its economics support scaling it, redesigning it or retiring it.

What MagicOrange does and does not do

MagicOrange names, measures and governs the cost. It does not route requests, choose models or stop an agent. The engineering team sets the stop condition. MagicOrange attributes, alerts, budgets and recharges, which is the finance side of the discipline, and the side nobody else holds across the whole estate.

See what sits inside the AI cost category

AI and agent workloads, the spend category page, lists what sits inside the category and how each cost moves through the model.

What it does today

AI spend already sits inside the model

Here is what the model already does, so you can test it.

AI cost and usage ingested as standard and not as an add-on.

Agents held as economic objects with an owner, a purpose and a result.

Speed of spend read as an anomaly signal beside the totals.

Cost per successful outcome, loaded across tokens, platform, data and people, in the same model as the other six spend categories.

AI agents need economic boundaries.

AI can remove friction from complex work. It should not remove financial accountability with it.

Book a demo

Reallocate capital to fund AI

Catch it before the invoice

Prove what AI returned