Forecast AI Spend Like a Variable Input
AI cost should not be forecast as 'last month plus growth'. It should be driven by customer count, usage frequency, workflow mix, model routing, success rates, and planned product changes.
- Driver-based forecasting
- Scenario planning
- Launch impact
- Provider price changes
- Budget ownership
The situation
AI spend can move quickly after a product launch, a routing change, a model upgrade, a viral customer workflow, or an increase in autonomous depth. A simple linear extrapolation often misses the reason spend is changing.
The naive solution
Take last month's provider bill and apply the revenue growth percentage.
The hidden constraint
Usage per customer and cost per workflow are not fixed. Product changes can alter both independently of customer growth.
The decision
Forecast from operational drivers. The model should make the assumptions visible so finance and engineering can challenge them together.
The implementation
A practical forecast can start with active customers × workflows per customer × successful-run cost, then add expected mix changes, retries/failures, free usage, internal usage, and planned premium features. I would build base, high-usage, and stress scenarios. Major releases should include an estimated cost effect in the launch plan, especially if they add images, voice, agents, longer context, or premium models.
The tradeoff
Driver models require maintenance as behaviour changes. That is still more useful than precise-looking forecasts built on assumptions nobody can inspect.
The evidence I would look for
Forecast accuracy should improve not only in total spend but in the team's ability to explain variance: more users, more runs per user, more expensive routing, lower cache rate, higher retry rate, or a pricing change.
The reusable lesson
Forecast the behaviours that cause AI spend, not the invoice line that appears after them.
The consultant's questions
- What operational assumptions drive the AI budget?
- How does the forecast change if usage per customer doubles?
- What launches in the next quarter materially change cost per workflow?
- Who owns the response when actual spend exceeds the forecast: finance, product, engineering, or all three?