Most 2027 AI budgets are built from list prices and seat counts, so they will miss the real bill, which depends on volume, work per task, model mix and repeats. It is for CFOs and technology leaders, who can use five controls and a 90-day plan to reforecast.
This paper establishes that 2027 AI budgets are larger versions of a number that was already wrong, because AI is bought by consumption and finance sees the bill rather than the meter. It sets out five finance controls and a 90-day plan to a first reforecast built from drivers.
Why are AI budgets going wrong?
In a CloudZero survey of 252 senior US finance leaders, 87% are raising their AI budget for fiscal 2027, and 61% say they already spend more on AI than they can justify. Futurum found 46.9% of enterprises over budget on AI in the second half of 2026 and 5.6% under. Among the over-budget organizations, 47.6% asked for supplemental funding and about one in six reduced or paused scope.
The paper breaks an AI bill into four drivers: volume, work per task, model mix, and repeats and checks. Price is a fifth factor and the least reliable guide. A research preprint, not peer reviewed, found that in 32% of model-pair comparisons the model with the lower list price cost more in total, by up to 28 times.
Where does finance lose sight of AI spend?
Deloitte's third-quarter 2026 CFO Signals survey found 53% of CFOs mainly review AI costs after the fact, through bills and reports. 74% said pausing over-limit usage would take several business days to several weeks. CloudZero found 54% of finance teams sit on the two lowest rungs of visibility, a total or a rough category split, and only 12% can allocate finely enough to act within the quarter.
What are the five controls?
- A unit: cost per completed task, defined for each workload that matters.
- An allocation: every dollar tagged to a workload, process and cost center when incurred.
- A driver-based forecast: a range built from the four drivers, not a single number.
- A limit and a fast stop: thresholds, alerts and a way to slow or pause usage in hours.
- A link to outcomes: spend reported next to what it produced.
The paper also gives an illustrative worked variance, using invented figures, and a plan of three 30-day phases: find what is being paid for, build the driver model, then reforecast and switch on controls.
Key takeaways
- 87% of surveyed US finance leaders are raising their fiscal 2027 AI budget, while 61% say they already spend more on AI than they can justify (CloudZero, 252 respondents).
- 46.9% of enterprises were over budget on AI in the second half of 2026 and 5.6% were under, according to Futurum Research.
- 53% of CFOs say AI usage costs are mainly reviewed after the fact, and 74% say pausing over-limit usage would take days to weeks (Deloitte).
- In 32% of model-pair comparisons in one preprint, the model with the lower list price cost more in total, by up to 28 times.
- Cost per completed task is the planning unit that links AI spend to something the business recognizes, counting tasks that finished correctly, not tasks started.
- Five controls are enough: a unit, an allocation, a driver-based forecast, a limit with a fast stop, and a link to outcomes.
What is in the full paper
- Executive summary
- Where budgets stand: Budgets rose without a model behind them
- The mechanics: Why AI spend does not behave like software
- The blind spot: Finance sees the bill, not the meter
- The framework: Five controls for consumption-priced AI
- The method: Reading the variance
- The plan: The first reforecast, a 90-day plan
- ETT Group: How ETT Group helps
- Next steps
The full paper has the figures, the sources and the working.
Download the PDF
