OpenAI pitches AI cost controls as agent use raises enterprise spend

OpenAI published a five-step framework for enterprise AI cost management on Tuesday, arguing that companies need tighter visibility into usage, outcomes and risk as agentic systems consume more capacity. The company did not disclose customer spending data or benchmarks, but its guidance reflects a growing problem for CIOs: lower token prices have not eliminated budget pressure as AI use spreads across the organization.

AI vendors have an interest in making enterprise deployments feel manageable, and OpenAI’s recommendations should be read in that context. Still, the cost issue is real. A Flexera study found roughly three-fifths of IT professionals said AI overspend has increased, while more than two-thirds said they lacked visibility into AI software usage.

The shift toward agentic AI raises the stakes because workflows can involve repeated model calls, tool use and retries before a business task is completed. That makes list prices for individual models an incomplete way to assess cost. OpenAI said companies should measure whether AI work creates value through metrics such as time saved, better decision-making and completed tasks.

OpenAI’s proposed cost playbook

OpenAI recommended that IT leaders start by building a clear view of who is using AI, which products and models are in use, how much capacity is being consumed and what work the usage supports. The company said rising bills are difficult to interpret without that baseline.

The second step is to judge models by the cost of achieving an accepted result, rather than by the sticker price of a model call. OpenAI said cheaper models can produce poor outputs or trigger additional attempts, which can increase total token consumption. For customer support, the relevant unit might be a resolved case. For engineering, it could be a tested change that clears review, according to the company.

OpenAI also said governance should define which AI work is allowed to scale. That includes setting limits on the data, context, applications and tools a large language model can access, as well as the actions it can take. The company said higher-risk steps should have defined approvers, and additional capacity should be granted based on workflow value.

The remaining recommendations are to manage AI spending as a portfolio and to fit the product, capacity and support model to demand once a workflow has shown value. OpenAI said stronger candidates for investment are repeated workflows with meaningful volume, clear ownership and measurable quality, risk and business value.

C-suite confidence is uneven

The guidance lands as enterprise leaders are still divided on AI returns. Protiviti found CIOs and IT decision-makers reported more confidence than CEOs and board members that AI can drive revenue growth. That gap matters for budget approvals, especially when pilot programs become ongoing consumption lines.

Some companies are already treating AI cost management as a formal operating discipline. Prudential Financial and Shutterstock have implemented stricter approaches, according to CIO Dive. Shutterstock CTO and CISO Courtney Totten said at FinOps X 2026 that understanding AI costs is no longer optional and has become foundational to business strategy.

The practical issue for operators is that AI spend is less predictable than traditional SaaS licensing when usage-based pricing, multiple models and agentic task loops are involved. OpenAI’s framework does not solve that by itself, and the company did not provide universal thresholds for acceptable cost per workflow. It does, however, point to where enterprise buyers are likely to push vendors next: more transparent metering, outcome-level reporting and controls that procurement, security and engineering teams can all use.

This story draws on original reporting from CIO Dive.