The new economics of AI cost management

For two years, enterprise AI adoption focused on usage volume, but organizations now prioritize cost management to ensure spending aligns with revenue growth. AI cost management involves measuring, attributing, and optimizing spend across models, agents, and teams to maximize business value per dollar. As AI agents handle more autonomous tasks, managing costs for reasoning, tool calls, and underlying models becomes essential. Databricks’ Kevin Stumpf emphasizes treating AI cost as a system to manage rather than a line item to minimize.

The industry is transitioning from 'tokenmaxxing'—rewarding raw usage—to 'valuemaxxing,' which prioritizes outcomes over volume. Early leaderboards incentivized heavy token usage, leading to inefficient practices like looping agents that burn tokens without adding value. Stumpf argues the new benchmark should rank teams by return on investment, a harder metric to quantify but the only one that truly matters for scaling AI responsibly.

AI cost attribution answers two key questions: where spend originates and what value it delivers. Databricks provides built-in attribution by user, team, agent, model, and time period, while organizations can customize tracking using request-level tags for projects or features. This granular visibility allows leaders to identify cost drivers and allocate resources effectively, addressing growing concerns about AI spending outpacing revenue.

Context bloat in production systems occurs when agents receive excessive irrelevant information, increasing token burn and decision overhead. Stumpf recommends letting agents pull only necessary context on demand through tools or MCP servers, rather than pushing everything up front. For model selection, dynamic routing can reduce costs by assigning tasks to the least expensive model capable of delivering required quality, with roughly 5–10% of tasks needing frontier models.