Every AI bill is a surprise the first time. Burn forecasting is how you make sure it's a surprise exactly once.
Why dashboards aren't enough
A spend dashboard shows cumulative cost over time. By the time the curve looks steep, the money is gone — you've paid for the diagnosis. A forecast inverts this: given what an agent has spent so far today and this month, where does the trend line cross the monthly cap? If the answer is "on the 18th," you act on the 11th.
The simple model works
You don't need machine learning for useful forecasts. Two projections cover most agents:
- Linear burn: today's rate, extended. If an agent spent $2.10 in 8 hours, the month projects to roughly $2.10 × 3 × 30. Crude, fast, and accurate for steady workloads.
- Growth-aware: weight recent days more than old ones. Agents' usage usually ramps; a 7-day weighted trend catches the ramp a linear model misses.
The output should be three numbers: projected month-end spend, the date the monthly cap will be hit (if ever), and the daily budget the agent must stay under for the rest of the month to land on target.
Forecast-driven throttling
A forecast becomes a control when it drives automatic behavior. Set a projected-burn threshold: if the forecast says the agent will exceed its monthly cap, tighten the daily cap automatically — the remaining budget divided by remaining days, minus a safety margin. The agent slows down instead of hitting a wall, and the owner gets a notification explaining why.
Seasonality is real for agents too
Batch jobs run on Mondays, support agents spike after releases, research fleets run overnight. If your forecast ignores day-of-week patterns, Monday's batch will look like a catastrophe every week. Track per-weekday baselines for at least a month before letting forecasts trigger automatic throttles — otherwise the throttle becomes the incident.
Forecasts are estimates; caps are law
Say it plainly to stakeholders: the forecast is a planning aid, the spend cap is the enforcement. Forecasts can be wrong (a quiet week, a one-off batch). Caps can't be. Keep the two visibly separate in the UI — projected vs. allowed — so nobody mistakes a projection for a guarantee.