AI economics · AI and agent workloads

AI workloads appear on invoices all over the business, but who owns the total?

A single agent run can be billed by the hyperscaler, the model API, the vector database and the orchestration layer. This page lists every cost inside AI and agent workloads, and who owns each one.AI economics is the subject. AI and agent workloads is the spend category. AI Economics with MagicOrange is how you govern it.

What this category is

Every cost created by running AI

AI and agent workloads is every cost created by running AI in the enterprise: the tokens a model uses, the API that bills them, the GPU capacity underneath, the agents that decide how much work to do, and the AI built into software you already pay for.

The FinOps Foundation defines its AI technology category as addressing “the cost complexity, faster development cycle, spend unpredictability, and the need for greater policy and governance to support innovation”, with activities of allocation, forecasting and optimization “to align AI consumption, investment, and business value.”

FinOps Foundation, FinOps Framework, Technology Categories, 2026.

Where AI sits among the seven spend categories

MagicOrange sorts technology cost into seven categories, grouped as Consumed, Owned or Human.

Consumed

Public cloudAI and agent workloads(you are here)Software and SaaSData cloud platforms

Owned

Infrastructure and networkEnd user technology

Human

Labor and services

What this category is not

AI cost does not behave like cloud cost.

The billing unit has no resource ID and no tag. It exists only in an API call.
The same prompt does not cost the same twice.
Frontier models are replaced within months, so a three-year commitment has a six-month shelf life.
The heaviest users forecast least accurately, because every efficiency gain unlocks new demand.

Inside the category

Six kinds of cost sit inside it. Three sit beside it.

From the part everyone measures to the parts almost nobody does.

Cost	What it is	Where it is billed
Tokens and inference	Input, output and cached tokens, by model tier, vendor and use case.	The model provider, or the hyperscaler’s AI service.
Model API and platform fees	Provisioned throughput, model hosting, fine-tuning and training runs, gateway and orchestration platforms.	Model provider, hyperscaler, orchestration vendor.
GPU and accelerator capacity	Reserved, on-demand and spot capacity, idle and shared.	Hyperscaler, specialist GPU cloud, or the data center asset register.
Agent workloads	The chain one request becomes: model calls, tool calls, retrieval, retries, sandboxes, guardrails and orchestration.	Spread across all of the above, plus external APIs.
AI inside software	Copilots, AI tiers and add-ons on software the business already licenses.	The software vendor’s invoice, often through a business unit.
AI arriving outside IT	AI tools bought on expense lines and business-unit budgets.	Expense systems and the general ledger.

Beside it, on the same model

These belong to AI programs but are booked elsewhere. MagicOrange reads them alongside this one.

Cost	Category it sits in	Why it is not here
The cloud the model runs on: compute, storage and networking	Public cloud	It is cloud cost, normalized to the FOCUS standard, whether or not AI used it. Much of the AI bill sits here.
The lakehouse and warehouse load that AI drives	Data cloud platforms	The compute layer that quietly carries much of the AI bill.
The engineers who build agents and the people who review their exceptions	Labor and services	Part of the true cost of every AI program, and a cost pool that cuts across every category.

A meter measures electricity. It does not run the grid. A token measures AI. It does not run the economics.

How it works

AI goes through the same six steps as every other cost

MOTEF, the MagicOrange Technology Economics Framework, puts every spend category through the same six steps. This is what each one does for AI.

See

Token usage, model API, agent workload and GPU arrive as standard objects beside cloud, software, the ledger, timesheets and assets. Model, vendor and use case are classified as the data comes in, so AI spend can be attributed in days, not quarters. AI that arrives through business units and embedded software is picked up when it enters the ledger.

What carries itAI Economics, with Data Studio included

Price

Cost per inference, per model, per agent run and per successful outcome, loaded across tokens, platform, data and people. Token and AI infrastructure costs are recharged to the business unit that used them, through the Bill of IT.

What carries itEnterprise Allocation Modeling, AI Economics, IT Financial Management

Plan

Token budgets by team and use case, model-switching analysis and consumption scenarios, so a change of model or volume is a scenario you run in hours, not a rebuild.

What carries itBudgeting and Forecasting, AI Economics

Optimize

Which models, routes and workflows give the outcome at the lowest loaded cost, with blended cost per million tokens as the measure for routing.

What carries itAI Economics, FinOps

Control

Cost allocated daily, not monthly. Speed of spend watched as an anomaly signal beside the totals: tokens per minute, calls per workflow, retry rate, incomplete runs. Findings go to Jira and ServiceNow with an owner.

What carries itFinOps, AI Economics

Prove

Cost per outcome set against the value of the outcome, by program, and read against plan.

What carries itBudgeting and Forecasting, with Analytics and AI Studio included

A new category, not a rebuild

AI arrived as a new category, a new row in the same architecture, handled the same way as cloud, software, data platforms, infrastructure, devices and labor. The model needed no rebuild, so the cost of the next category is close to zero.

Can AI spend be capitalized?

The same spend can be expensed or capitalized. AI development cost follows existing accounting standards, not a new AI rule: IAS 38 under IFRS, and ASC 350-40 with ASU 2025-06 under US GAAP. Under both, you can capitalize only what you can evidence: the project, the work and the hours.

Where the model keeps the evidence

It sits in one place: project and development-stage tagging, timesheet hours by project from Time Capture and Resource Planning, AI usage as standard objects, and an OpEx and CapEx split with amortization modeled once an asset exists. The accounting policy is the customer’s. The platform applies it and shows its working.

Shown as what the model does, not as a customer result.

Questions answered

Eight questions about AI cost, answered from one model

What did AI cost us this month across the model bill, the cloud it ran on, the data platform and the people, and who owns each part?
How much of our AI spend never passed through IT or finance, and where did it arrive?
What does one inference cost, and what does one successful outcome cost fully loaded?
Which agents consumed what, across model calls, tool calls, retries and duration, and what did each produce?
What is our blended cost per million tokens, and is routing to cheaper models working?
If we switch this model, or volume doubles, what happens to the plan?
How much of the AI program’s cost is eligible to capitalize under our accounting policy, and where is the evidence?
Which AI programs are returning value above their cost, and which are running as activity?

What it does today

Every AI cost lands with an owner

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

AI cost and usage ingested as standard and classified by model, vendor and use case.

Attributed to an owner and recharged through the Bill of IT.

Planned as consumption scenarios and controlled daily.

Read as cost per inference and cost per successful outcome beside the other six spend categories.

How the capabilities connect

AI Economics governs AI cost. Five other capabilities share it.

The capability that governs AI cost

AI Economics, with six capabilities: token and model cost, agent and GPU cost, AI unit economics, AI ROI, shadow AI discovery and token budget scenarios.

The capabilities that share AI cost

FinOps for the cloud it runs on.
Enterprise Allocation Modeling and IT Financial Management for allocation and the Bill of IT.
Budgeting and Forecasting for the plan.
Time Capture and Resource Planning for the people.

Included with every capability

Data Studio ingests it, Analytics reports it, AI Studio answers questions over it.

The argument

The challenge

AI and agent economics (on the solutions page)

Built on the standards

FinOps for AIFOCUSTokenomics Foundation

All AI costs. In the same model as everything else.

Book a demo