How to manage AI investments in the agentic era

OpenAI has published guidance aimed squarely at the people signing off on enterprise AI budgets, and the message is blunt: the old ways of measuring AI spend do not work once agents enter the picture. Counting API calls or tracking token usage might have made sense when AI was a chatbot answering one question at a time. It makes a lot less sense when an autonomous agent is chaining together dozens of steps, calling tools, checking its own work, and running for minutes or hours to finish a task.

The core idea is to shift the metric entirely, toward something OpenAI calls useful work per dollar. Instead of asking how much a model call costs, the question becomes whether the output actually moved a business process forward, and at what total cost including compute, tool use, and retries. That reframing matters because agentic workflows are inherently more expensive per task than a single prompt-response exchange. An agent might explore several dead ends before landing on a correct answer, and all of that exploration shows up on the bill.

OpenAI's guidance pushes enterprises to get comfortable with efficiency as an ongoing discipline rather than a one-time optimization. That means monitoring which workflows are actually generating value versus which ones are just consuming compute for marginal gains, and being willing to kill or redesign the latter. It also means picking the right model tier for the right job, since running a frontier-scale model on a task a smaller, cheaper model could handle is exactly the kind of waste that erodes returns as agentic use scales up across an organization.

The timing here is not subtle. As more companies move from pilot projects to production deployments of AI agents, the costs scale in ways that traditional software budgets never had to account for, because compute is now a variable cost tied directly to how much reasoning an agent does. OpenAI is effectively telling enterprise buyers to treat agent spend like a portfolio, one where some workflows deserve heavier investment because they generate outsized value, and others should be trimmed or automated more cheaply. Whether companies actually build the measurement infrastructure to do this well, rather than just nodding along and continuing to overspend, is the open question.