An AI pilot can have an approved budget, an executive sponsor, and no agreed answer to one basic question: who can stop it spending money? That is an awkward gap to discover after the invoice arrives.
CFO.com reported on Deloitte’s third-quarter 2026 CFO Signals survey: 53% of responding CFOs said their organizations primarily review AI usage costs after the fact. When cost controls fail, 74% said pausing the related usage typically takes several business days to several weeks. Deloitte surveyed 200 CFOs at North American companies with annual revenue of at least $1 billion; these are findings about large organizations, not a census of every U.S. business.
The practical question of how CFOs control AI usage costs starts well before month-end. An invoice is excellent evidence that a cost happened. As an early-warning system, it’s a bit late.
Give each workflow an owner and a useful cost unit
A monthly AI bill cannot explain whether spending rose because more customers used a service, a new model costs more, or employees had to rerun poor results. Those are three different management decisions.
Start by defining the result each workflow is meant to produce: a resolved customer case, a reviewed invoice, or a forecast scenario that someone can actually use. Track model usage and vendor charges, but report the cost per accepted result, including retries. Add platform fees, supporting infrastructure, and human review when assessing the full economics. The FinOps Foundation’s unit-economics guidance makes the same useful distinction between a technical consumption metric and a business outcome.
Then assign accountability. Engineering should own reliable usage data and technical safeguards. The business owner should own output quality and whether extra volume is worthwhile. FP&A should own the spending forecast and explain the variance. One designated approver needs authority to change or suspend the plan. A committee can discuss the bill; it cannot be the person who turns something off.
Forecast AI usage costs from business drivers
For a usage-priced workflow, a workable starting model is completed tasks multiplied by average attempts per completed task, multiplied by the expected cost per attempt. Add the fixed license, infrastructure, and review costs separately. The exact calculation should follow the vendor’s real pricing terms, particularly where input, output, caching, and other services carry different charges.
Consider a hypothetical example. A workflow completes 50,000 useful tasks a month at $0.04 in variable AI cost per completed task: $2,000. Next month, volume reaches 80,000 and variable cost rises to $0.06 per completed task. Spending becomes $4,800, a 140% increase. These figures are illustrative and exclude fixed costs.
A $2,800 unfavorable variance is accurate. It is also incomplete. Did demand expand profitably? Did staff rerun more outputs? Was a more expensive model introduced without improving quality? Finance needs a volume-and-rate bridge, alongside error rates and the business benefit, before deciding whether the higher bill is good news or a problem.
Agree on intervention rules before the overrun
Each production workflow needs a spending envelope, an early-warning trigger, a decision trigger, and a named owner for exceptions. For example, a team might alert the owner when 70% of its monthly budget has been consumed and require a fresh forecast and approval when the projected month-end run rate exceeds plan. Those are illustrative policy thresholds, not universal benchmarks.
Where vendor controls allow it, engineering can enforce limits or switch noncritical work to a cheaper configuration. A customer-facing process may need a fallback rather than an automatic shutdown. Finance should know which workloads can pause safely and which require a documented exception. The authority to approve that exception should be established before the alert, not discovered through an urgent email chain.
Review a short weekly scorecard: spend to date, expected month-end spend, cost per accepted outcome, retry or rework rate, service quality, and a measurable business result. A workload that exceeds its budget because profitable demand increased may deserve more capacity. One that stays within budget while the useful-output rate collapses may deserve less. The green cell in the budget is not the verdict.
The cost of doing nothing is the time between an avoidable problem starting and somebody having enough information and authority to act. For a CFO, that is a forecast design problem as much as a technology problem.
The Schlott Company helps finance teams build driver-based forecasts, scenario models, and management reporting that connect changing costs to actual decisions. If AI spending is moving faster than the process for evaluating it, contact us to discuss the forecast and controls your team needs.