How to Forecast AI Demand and Set Budgets That Hold
An AI budget fails in a particular way. Unit prices fall, so the budget looks safe, while usage climbs faster than anyone planned, so the bill blows through it anyway. Agent-driven work makes this worse, because a single request can cost many times what the chat it replaced did, and two identical runs can differ wildly. This guide shows you how to build forecasts and budgets that survive that reality instead of being wrong on arrival.
Developing
Start here. Build the foundation.- 1
Before you project anything forward, establish what the workload costs today, split into price per unit and units consumed. Pull the baseline from real usage over a full cycle, not a list price times a guessed request count, because business cases built on sticker prices are the most common reason an AI investment evaporates in production. You have a baseline when you can state both halves of it from data.
- 2
When a workload runs an agent that loops or calls tools, give the forecast a low, expected, and high case instead of a single number, and size the budget against the high case or say out loud that you are choosing not to. At this stage an honest range beats a precise guess. You are done when the forecast shows its spread rather than hiding it.
Proficient
Build consistency and rhythm.- 3
Before asking for approval, walk finance through what drives the spend and which assumptions move it, so they understand the mechanism and not just the total. You know it worked when they can restate the cost drivers back to you and challenge an assumption you did not point at. A number with no mechanism behind it gets cut like a guess.
- 4
When a vendor offers a discount for committed volume, size the commitment to the floor of demand you have seen hold across several cycles, not the forecast ceiling, and report any unused commitment honestly rather than burying it. The signal you got it right is that the committed portion stays fully used cycle after cycle.
Mastered
Operate at the highest level.- 5
Create the forecasting model other teams fill in with their own workloads to produce a defensible budget, one that treats falling price and rising volume as separate inputs rather than a single blended trend. You know it works when at least two teams outside yours use it in a planning cycle without your help. A model only you can run is a personal spreadsheet, not a standard.
Common Pitfalls
Avoid the common failure modes.- Building the business case on the vendor's sticker price with no allowance for usage at scale, so the real bill lands far above plan.
- Presenting agent-driven work as a single number, which quietly hands everyone the risk of a run that costs ten times the estimate.
- Over-committing to a reservation to capture a headline discount, turning spend you could have cut into a fixed cost you cannot.