The 2024 AI spend forecast for most organizations was wrong in the same direction: too low. The methodology that produced those forecasts had a structural flaw — it treated AI spend as a linear extrapolation of recent costs without modeling the non-linear adoption dynamics that drive actual AI spend growth.
A forecast that projects next quarter's AI spend as 10 percent above this quarter's will be accurate during a slow adoption phase. It will fail dramatically during an acceleration phase, which is where most organizations are right now. The forecast fails not because the math is wrong but because the model uses the wrong variables.
Why linear extrapolation fails
AI spend has two growth drivers that compound differently. The first is adoption breadth: as more employees and more use cases come online, spend grows. The second is adoption depth: as existing users discover more ways to use AI tools, and as agentic systems and automated workflows scale with business activity, spend per existing user grows.
A linear extrapolation captures neither driver accurately. It assumes that whatever rate of growth you have seen recently will continue at roughly the same rate. But adoption breadth tends to grow in steps — a new tool gets approved, an enthusiastic team adopts it, and then usage diffuses across the organization in a pattern that is faster at the beginning and slower as the population saturates. And adoption depth is notoriously hard to project because it depends on how teams evolve their AI workflows, which is hard to forecast without direct input from team leads.
The input variables that produce accurate forecasts
A forecast that survives contact with reality needs to model three things the linear approach ignores.
User adoption rate by team. Ask each team lead two questions: what percentage of your team is actively using AI tools today, and what percentage do you expect to be actively using them in six months? The difference is the adoption growth driver. For teams that are at 90 percent adoption already, this driver is nearly exhausted. For teams at 30 percent, it could drive 2x to 3x spend growth even with no changes to tools or pricing.
Per-user spend intensity changes. Separately from whether new users are adopting, ask whether existing users are using AI more intensively over time. The proxy metric is average spend per active user per month, trended over the last six months. If this number is growing — because users are discovering new use cases, because workflows are becoming more AI-dependent, because agentic systems are scaling — that trend needs to be modeled separately from adoption growth. The two compound.
Pipeline of new use cases. A forecast built only on historical data will miss the spend impact of use cases that are in development but not yet in production. Ask engineering leadership: what AI-powered features or automations are in the development pipeline for the next six months? For each, estimate the expected API call volume at steady state and the expected ramp timeline. Add this pipeline spend to the baseline forecast with an appropriate probability weight.
Building the model
The model structure is simple: (current active users x current average spend per user x adoption growth factor x intensity growth factor) + pipeline spend = projected spend.
The growth factors are the inputs that require judgment. Adoption growth factor is 1.0 if adoption is saturated, and higher — potentially 1.5 to 2.0 — if significant portions of the potential user population have not yet adopted. Intensity growth factor reflects whether per-user spend is growing and at what rate. A team whose average spend per active user has grown 15 percent month-over-month for three consecutive months has a higher intensity growth factor than one whose per-user spend is flat.
Run the model monthly. The monthly update is not a large exercise — it is updating the inputs based on current data and observing whether the model's projections from last month were accurate. When projections are consistently off in one direction, the growth factor assumptions need calibration.
The forecast conversation with finance
A forecast is most useful when it is the beginning of a conversation, not a number on a slide. When you present the AI spend forecast to finance, present it with a range and a set of explicit assumptions.
The range acknowledges that the growth factors are estimates. A low-growth scenario where adoption growth slows and intensity stays flat, a base scenario with current trend assumptions, and a high-growth scenario where a major new use case comes online ahead of schedule — all three should be on the table.
The explicit assumptions let finance pressure-test the model. If they believe adoption will be slower than you modeled, they can say so. If they believe a specific use case will scale faster, that input goes into the high-growth scenario. The conversation produces a forecast that finance owns alongside engineering, which means it is actually used in planning rather than filed and forgotten.
Most AI spend surprises are not surprises at all in retrospect. The drivers were visible. The forecast methodology just wasn't built to see them.