Finance teams did not choose to own AI spend; it landed on them. It arrived as a fast-growing, unpredictable line item that behaves nothing like the software costs they know how to forecast. Tracking it well is less about new tools than about applying financial discipline to a category that resists it.
Why AI spend resists standard tracking
Traditional software is seat-based and predictable. AI spend is consumption-based and volatile, scaling with usage rather than headcount and swinging month to month with feature launches and customer behavior. FP&A models built for fixed subscriptions break when applied to a line that can double in a month for engineering reasons no one flagged.
Separate the two kinds of AI spend
The first move is to split AI spend into its two behaviors. Subscription and seat spend is predictable and forecasts like any other software line. Consumption spend, from APIs and product features, scales with activity and needs a usage-based model. Tracking them together as one number hides both; separating them makes each forecastable.
Attribution is the finance foundation
FP&A cannot allocate what it cannot attribute. Insist that AI spend map to teams, projects, or cost centers through API key structure and seat assignment. Once every dollar has an owner, AI spend fits into the same allocation and accountability framework as every other cost, and monthly close stops being a reconstruction exercise.
Forecast from drivers, not from history
A linear extrapolation of last month's AI spend will be wrong, because the drivers are non-linear. Forecast instead from the underlying drivers: user adoption rate, per-user intensity, and the pipeline of new AI features heading to production. Tracking those drivers is what makes the forecast hold up in a board meeting.
Finance can bring AI spend under control the same way it brought every other volatile category under control: attribute it, separate its behaviors, and forecast from drivers. The category is new; the discipline is not.