AI ECONOMICS · JUN 9, 2026
Why AI Agent Spend Is So Hard to Forecast
Teams rolling out AI coding agents are burning through annual budgets in months. The problem isn't carelessness; agent spend is genuinely hard to predict. Here's how to forecast it.
Table of Contents
- Why agent costs are hard to predict
- What a useful forecast looks like
- The takeaway
This year brought reports of large engineering organizations rolling out AI coding agents to thousands of developers and exhausting a full year's AI budget within a few months. Those teams weren't careless. They ran into a problem every team running agents eventually meets: agent spend doesn't behave like a normal software line item.
Why agent costs are hard to predict
- Usage grows unevenly as people discover new workflows.
- Token consumption varies wildly from task to task and model to model.
- Traditional budgeting assumes steady, per-seat costs, and agents don't follow that shape.
What a useful forecast looks like
The question that matters is simple: where will this land in six months, and what can we change before then? Answering it takes a forecasting layer that:
- reads usage data straight from a live agent deployment or a data warehouse,
- returns a calibrated, explainable forecast in seconds without training a model first, and
- lets teams run what-if scenarios in plain English.
The takeaway
If you're scaling AI agents, forecast the bill before it forecasts itself.
