How to budget for AI when nobody knows next quarter's usage

In brief

Split AI spend into five lines: plan the fixed ones, forecast consumption against a range and size contingency to it. A worked budget and the caps that work.

Steve Elliott

Steve Elliott · · 6 min read


On this page
  1. Why does a single AI budget line break?
  2. Which AI costs should you plan, and which should you forecast?
  3. What does an illustrative AI budget look like?
  4. How much contingency should an AI budget hold?
  5. Should high-case demand go into capacity commitments?
  6. Which spend controls actually stop AI spend?
  7. Who should approve what?
  8. Sources

In our view, an AI budget should be five lines, not one: seat licenses, API consumption, capacity commitments, shared infrastructure and people. Plan the four that contracts and headcount fix, forecast consumption every month against a range, and hold one contingency centrally, sized to how wide that range is.

Only one of those lines is mainly driven by usage, and folding it into the others hides it. Below: a worked illustrative budget, a contingency method, and which provider spend controls stop traffic and which only send an email.

Why does a single AI budget line break?

Because it mixes costs that behave differently. A seat license is price times headcount, fixed at contract. A capacity reservation is billed whether used or not. Token consumption moves with adoption, prompt design and model choice, and can, for example, double in a quarter without anyone signing anything.

The FinOps Foundation's FinOps for AI guidance says predictability "is generally lower, especially for the Crawl and Walk phases," that higher cost volatility makes trend-based forecasting harder, and that budgets need more frequent revision. In its State of FinOps 2026, released February 19, 2026, 98% of respondents said their FinOps practice now covers AI spend, up from 63% in 2025 and 31% in 2024.

An AI budget with one line can't tell a CFO whether an overrun came from success (more use of a useful service) or waste (the same work costing more per task).

Which AI costs should you plan, and which should you forecast?

Plan what a contract or a hiring decision fixes. Forecast what usage drives.

LineWhat drives itTreatmentMain risk
Seat licenses (copilots, chat assistants)Seats times pricePlan annuallySeats bought and not used
API consumptionTasks times tokens per task times priceForecast monthly against a rangeOverrun from adoption or design
Capacity commitments (provisioned throughput, GPU reservations)Term and size of the commitmentPlan for the termPaying for idle capacity
Shared infrastructure (AI gateway, observability, vector databases, evaluation runs)Mostly fixed, partly usage-basedPlan the fixed part, forecast the restCosts nobody owns
People (engineering, evaluation, change)Headcount and contractorsPlan annuallyHiring slower or faster than demand

The FinOps Foundation's paper on forecasting AI services costs recommends forecasting key domains such as retrieval-augmented generation, fine-tuning and training as quantity times rate, and a weekly or monthly forecast cadence because AI costs vary so much. That cadence belongs on the consumption line, carried in a rolling forecast; the other four sit in the annual plan. See annual budget vs rolling forecast, forecasting volatile cloud spend and, on labor versus tokens, AI initiative costs: tokens vs labor.

What does an illustrative AI budget look like?

The figures below are made up to show the mechanics. Every price is an assumption, not a vendor quote.

LineBasis (illustrative)Annual amount
Seat licenses1,500 seats at an assumed $30 a seat a month, times 12$540,000
API consumption, most likely400,000 tasks a month at an assumed $0.25 a task, times 12$1,200,000
Capacity commitmentOne reservation at an assumed $15,000 a month, times 12$180,000
Shared infrastructureGateway, observability, vector database, evaluation runs$450,000
People6 full-time equivalents at an assumed $210,000 fully loaded$1,260,000
Subtotal$3,630,000
Contingency, held centrallySized from the consumption range (next section)$560,000
Total$4,190,000

Consumption is budgeted as tasks times cost per task, not a token total, so the monthly reforecast shows which of the two moved. Contingency is a separate line held by finance; we think that works better than spreading it across teams, where it tends to get spent.

How much contingency should an AI budget hold?

Enough to cover the plausible high case on the lines that can move, not a flat 10% on everything. One method in the US Government Accountability Office's Cost Estimating and Assessment Guide (GAO-20-195G) builds a range of possible costs from low, most likely and high estimates, then sets contingency by choosing a confidence level on that range. In its worked example, the sum of most likely costs had only about a 15% chance of holding.

Apply the same idea to the illustrative consumption line, with three estimates:

  • Low: $700,000. Adoption is slower than planned.
  • Most likely: $1,200,000. The plan.
  • High: $2,400,000. Adoption doubles, or agent workflows use far more tokens per task than tested.

The range is skewed: $1.2 million above plan, $500,000 below. Treated as a triangular distribution, the simplest three-point model, our arithmetic gives:

  • Chance the most likely figure holds: (1.2 − 0.7) ÷ (2.4 − 0.7) = about 29%.
  • Expected cost: (0.7 + 1.2 + 2.4) ÷ 3 = about $1.43 million, already above the plan.
  • 80th percentile: 2.4 − √(0.2 × 1.7 × 1.2) = about $1.76 million.

Contingency at 80% confidence is therefore about $1.76 million minus $1.2 million: roughly $560,000, or 47% of the consumption line and 15% of the subtotal. In our view, the fixed lines need little or none, and adding each line's 80th percentile together would overstate the total, because lines rarely all run high at once.

We think the release rule matters as much as the size. Release contingency when the monthly reforecast shows consumption above plan and cost per task is at or below target: that is growth. If cost per task is the thing that rose, the answer is an engineering fix, not more money.

Should high-case demand go into capacity commitments?

In our view, no. Commit to the baseline you have measured, and let the high case run at pay-as-you-go out of contingency. The FinOps Foundation's forecasting paper warns against "committing too much for too long" without allowing for future advances and optimizations. Its paper on the effect of optimization on AI forecasting, updated March 17, 2026, says costs tend to decrease while quality and performance make significant jumps every 6 to 9 months. A one-year commitment sized to the high case may be too big and on the wrong model by month nine. Break-even utilization for each provider's capacity product is in AI capacity reservations vs cloud commitments.

Which spend controls actually stop AI spend?

Not all of them, and no provider documents a spend cap that acts instantly. As of November 2026:

ControlScopeStops traffic?Caveat in the provider's docs
OpenAI spend limitsOrganization or project, monthlyYes, if "Enforce a hard limit" is on; requests then return 429 errors. Spend alerts alone don't cap trafficNot instantaneous; recorded spend can slightly exceed the limit
Anthropic spend limitsConsole: organization or workspace, monthly. Claude Enterprise: per-member limits on usage beyond the planCaps monthly spend; the docs don't state what happens to requests at the limitWorkspace limits can't raise organization limits; a Default Workspace limit may not count Console playground, Managed Agents or some Claude Code usage
Google Cloud spend cap budgetsOne project and one eligible service, monthly; covers the Gemini API and Gemini Enterprise Agent Platform (formerly Vertex AI), including usage covered by committed use discountsYes: pauses that service in that project for the rest of the periodNot instantaneous; persistent resources keep billing; Google advises setting the cap slightly below your real limit
AWS Budgets actionsAccount, via IAM policy or service control policyOnly if you configure an action, automatic or after approvalBudget data updates up to three times a day
Azure budgetsScopes including subscription and resource groupNo: "your consumption isn't stopped"; action groups can trigger automation at subscription and resource group scopeBudgets are evaluated every 24 hours
Azure OpenAI quotaTokens per minute, per subscription, region, model and deployment typeYes, as a rate limit: requests beyond it get 429 errorsA throughput limit, not a spend limit

In our view, a cap that lags by hours, as AWS and Azure budgets do, is a backstop, not a budget, and a hard cap on a customer-facing service is an outage scheduled in advance. Microsoft's quota page notes that 429 errors in development or testing may be intentional cost guardrails.

Who should approve what?

We suggest guardrails that tighten as the money at stake grows. The amounts are illustrative.

  1. Every project or workspace gets its own key and a monthly limit at creation. Sandboxes and pilots get a hard limit, for example $1,000 a month. The FinOps Foundation paper recommends fixed budgets for proofs of concept and sandboxes, with real-time alerts as thresholds approach.
  2. Production services get alerts and rate limits, not hard spend caps. Alert the owner at 50%, 80% and 100% of the monthly figure, and use tokens-per-minute limits to stop runaway loops.
  3. Raising a limit above $10,000 a month needs the budget owner. The request states expected tasks and cost per task.
  4. Above $50,000 a month, or any capacity commitment, needs finance. Commitments need a measured baseline and a break-even calculation.
  5. Reforecast consumption monthly and report it against the range, not just the plan. A month at the high end of the range is information; a month outside it calls for a review.

The paper's success indicator is forecasts within 5% of actuals per month. We think that is fair for a mature service and hard in a first year, which is what contingency is for.

A budget only explains an overrun if each line has an owner and each dollar can be traced to the service and decision behind it, which is the case for keeping AI funding inside the same operating model as the rest of technology spend.

Sources

Frequently asked questions

Should AI spend be one budget line or several?

In our view, several: seat licenses, API consumption, capacity commitments, shared infrastructure and people. They behave differently. Seats and people are fixed by contract and headcount, commitments are billed whether used or not, and consumption moves with adoption and design. One line can't show whether an overrun came from more use or from higher cost per task, so finance can't tell success from waste.

How much contingency should an AI budget include?

Size it from the range on the lines that can move, mainly API consumption, rather than a flat percentage. Estimate low, most likely and high cases, then budget to a chosen confidence level, as one method in the GAO cost guide does. In our illustrative example, an 80% confidence level needs about $560,000 on a $1.2 million consumption plan, about 15% of the pre-contingency subtotal.

Do cloud budgets stop AI spending when they are exceeded?

Not by default. As of November 2026, Azure budgets send notifications without stopping consumption, and AWS Budgets restrict spend only through configured actions such as IAM or service control policies. Google Cloud spend cap budgets pause eligible services, including the Gemini API. OpenAI and Anthropic offer monthly spend limits by project or workspace. OpenAI and Google both say enforcement is not instantaneous.

How often should AI spend be reforecast?

Monthly at least for API consumption. The FinOps Foundation recommends a weekly or monthly cadence because AI costs vary so much; we would use weekly during rapid adoption. Seat licenses, capacity commitments and people can stay in the annual plan and change only when a contract or headcount decision does. Report consumption against a range so a high month is read in context.

Free playbook

Run the enterprise from any AI.

Read The Headless Enterprise
Steve Elliott

Steve Elliott

Published November 22, 2026

More on AI Economics

Tokens, inference cost, cost per task.