Usage-priced AI breaks annual budgets because the cost line moves with behavior you do not control: how many requests users send, how long the prompts and answers run, and which model handles each one. A fixed budget assumes spend follows headcount or contracts. Token-metered spend follows product adoption and engineering choices. The fix is to budget per workflow, split committed from on-demand capacity, and decide the accounting treatment before the first invoice arrives.
Why consumption pricing does not fit an annual plan
A traditional software budget is a list of contracts with a renewal date. AI spend is closer to a utility bill. Microsoft's FinOps guidance for Azure OpenAI states that the service charges based on token usage, and that cost depends on the model you choose and how long your prompts are.
That creates three forecasting problems. Demand is uncertain, because a feature that works can double its traffic in a quarter. Unit cost is uncertain, because a prompt that grows from a short question to a long document changes the bill for the same request. And the mix is uncertain, because engineers can change the model behind a feature in a single release.
The FinOps Foundation's paper on FinOps for AI makes a related point about scope. It says AI spend can span model provider APIs, cloud AI services, SaaS and developer tools, GPU providers, and owned data center infrastructure. A finance team that reads only the cloud invoice will miss part of the picture, and a budget owner who sees only one channel cannot explain a variance.
A simple way to see the problem is to run the budget as a range. Consider a company that plans one workflow at an expected monthly volume and then re-runs the model at double the volume and at double the average prompt length. The two sensitivities multiply, so a fourfold swing in spend is possible without any change in price. A budget that shows only the expected case hides that swing from the board.
Build unit economics per workflow, not per token
Cost per token is a useful normalizing metric. Microsoft's toolkit guidance calculates it as billed cost divided by consumed quantity, using FOCUS-formatted billing data, and suggests splitting cost by meter so that input and output tokens are visible by model. That gives finance a clean series to trend.
It is not enough for decisions. The FinOps Foundation paper recommends tracking cost per successful outcome, reported as a range, over raw token cost, and tying unit economics to results the business owns. A claims-triage workflow, a contract-review workflow and a customer-support workflow each have a different cost per completed task, a different quality bar and a different value per task.
A workable unit-economics sheet for each workflow has five lines:
- Volume: completed tasks per month, with a low, expected and high case.
- Cost per attempt: model fees, retrieval, tool calls and any human review time.
- Success rate: the share of attempts that meet the quality bar without rework.
- Cost per successful outcome: cost per attempt divided by success rate, shown as a range.
- Value per outcome: owned by the business unit, not by finance or engineering.
Retries and failed attempts matter here. If a workflow succeeds on 70 percent of attempts, the other 30 percent still cost money. Counting only successful calls understates the true unit cost, which is why the paper's emphasis on outcomes rather than tokens matters for a CFO.
Committed versus on-demand capacity
Providers sell the same model capacity in different structures, and each one moves risk between you and the vendor.
On-demand, per-token pricing carries no commitment. You pay for what you use, and the cost of a bad forecast is a surprise invoice. It is the right structure for pilots and for workloads whose volume you cannot predict.
Provisioned capacity works the other way. Microsoft's documentation for provisioned throughput says you are billed hourly for the provisioned throughput units you deploy, rather than the tokens consumed, and that you pay for the full deployed count regardless of actual utilization. Reservations for those units run for a one-month or one-year term at a discounted effective rate. Units deployed beyond the reserved quantity are billed at the hourly rate, and the documentation notes that cancellations or exchanges might incur fees. It also warns that a reservation does not guarantee capacity, so deployments should be created first and the reservation bought afterward.
Providers also sell discounts for flexibility on your side. OpenAI's Batch API documentation describes batch jobs costing 50 percent less than synchronous calls, with a 24-hour completion window, and names evaluations and large classification jobs as typical uses. Anthropic's documentation describes prompt caching with a premium to write a cache entry, a much lower price to read one, and a choice between a five-minute and a one-hour lifetime.
The practical rule follows from these structures. Commit only to the steady base load that you have measured over several months. Put predictable but not urgent work into batch queues. Keep spiky and experimental traffic on demand. Review each commitment against the unit-economics sheet before renewal, and treat reservation utilization as a finance metric, because Microsoft's documentation notes that utilization below 100 percent means reserved units are going unused.
Capitalize or expense
Accounting treatment is where the budget meets the financial statements, and the answer depends on what you are actually paying for.
Per-token fees for a hosted model look like consumption of a service, and we would expect them to be expensed as incurred. A commitment structure, such as a one-year reservation, needs a separate look at what is being acquired and over what period. Settle both questions with your auditors before you sign.
Building your own AI-enabled software is a different question. Under US GAAP, the FASB issued Accounting Standards Update 2025-06 in September 2025, covering internal-use software under Subtopic 350-40. It removes references to development project stages. Under it, an entity starts capitalizing software costs when management has authorized and committed to funding the project, and it is probable that the project will be completed and the software used for its intended function.
The update adds a test for significant development uncertainty. Two factors count: whether the software has novel or unproven features whose uncertainty has not been resolved through coding and testing, and whether the entity has determined what the software must do, including whether it keeps substantially revising its performance requirements. An AI feature can fail both tests in its early months, because model behavior is unproven until it has been tested against real data. That argues for a documented decision record at each stage, not a single capitalization call at kickoff.
The effective date is annual periods beginning after December 15, 2027, with early adoption permitted and a choice of prospective, modified or retrospective transition. The FASB also says it expects capitalization to decrease for software delivered through a cloud computing arrangement. Note that the update does not affect software to be sold, leased or marketed, which falls under Subtopic 985-20.
Under IFRS, IAS 38 draws the line differently. The IFRS Foundation's summary states that research expenditure is recognized as an expense, while development expenditure that meets specified criteria is recognized as the cost of an intangible asset. Prototyping and evaluating an AI use case sits near the research end. Finance teams should document when a workflow moves from exploration to development, because that date drives the accounting.
FinOps practices that apply to AI
The FinOps Foundation paper says most existing FinOps practices carry over, and that tokens, volatile pricing and new stakeholders are what change. Its sequence is a useful checklist: see and allocate spend first, then right-size models and use tokens efficiently, and commit to discounts once usage is stable.
Allocation depends on tagging. Microsoft's guidance suggests tags such as cost center, environment and feature, so that spend can be attributed to teams or products for showback or chargeback. If an application calls a model through a shared gateway, require a workflow identifier on every request, or the cost will land in one unallocated bucket.
Model choice is a financial decision. The paper advises choosing the least expensive model that meets the task's quality bar. That requires an evaluation set for each workflow, so that a cheaper model can be tested against the current one before anyone switches.
The paper also describes a Crawl, Walk, Run progression, from pilots to integrated production use to core business processes. Controls should match the stage. A pilot needs a spending cap and an owner. A production workflow needs unit-cost reporting, alerting on variance and a quarterly review of its commitment level.
Where to start
Start with measurement before negotiation. These first steps fit inside a quarter:
- List every channel that carries AI spend: provider APIs, cloud AI services, SaaS tools, GPU capacity and internal infrastructure.
- Require a workflow tag on every model request and map billing data to a common schema such as FOCUS.
- Build the five-line unit-economics sheet for your three largest workflows, with low, expected and high volume cases.
- Move non-urgent, high-volume work to batch processing and test prompt caching on repeated context.
- Size any commitment at the steady base load you have measured, and review utilization monthly.
- Ask your controller and auditors for a written view on the treatment of per-token fees, reservations and internal build costs, and on the transition date for ASU 2025-06.
Start with the workflow ledger, because a budget model built on that ledger can answer the board's question of what each business outcome costs.
References
- FinOps Foundation: FinOps for AI
- Microsoft Tech Community: Managing Azure OpenAI costs with the FinOps toolkit and FOCUS
- Microsoft Learn: Provisioned throughput billing and cost management
- OpenAI: Batch API guide
- Anthropic: Prompt caching documentation
- FASB: Accounting Standards Update 2025-06, Internal-Use Software (Subtopic 350-40)
- IFRS Foundation: IAS 38 Intangible Assets





