Five Ways to Manage Enterprise AI Spending by Value

OpenAI argues that enterprise AI spending should be managed around the value of completed work rather than the price of individual tokens. Although the company says its price per million tokens fell by 97% between GPT-4 and GPT-5.4, lower unit prices do not automatically translate into better economics. It also cites GPT-5.6 as an example of efficiency gains, saying the model performs better on the Artificial Analysis Coding Agent Index while using 54% fewer output tokens and taking 57% less time per task. The central recommendation is to measure useful work per dollar: completed tasks, saved time, improved decisions, and workflows that can be scaled.

This shift becomes more important as employees move beyond short chat exchanges and start using AI for long-running, multi-step workflows. Enterprise administrators need to understand who is using AI, which products and models they select, how much capacity they consume, and what business activities that consumption supports. A rising bill is ambiguous without this context: it may indicate waste, healthy experimentation, growing adoption, or the emergence of a recurring process that has become essential to the business.

OpenAI presents updated analytics and spending controls in the ChatGPT Admin Console as tools for building that visibility. Administrators can examine adoption, credit consumption, and spending by user, product, and model, as well as follow trends and identify recurring patterns. The analysis should operate at several levels. At the workspace level, leaders can compare adoption with total spending. At the team and individual level, they can find areas of increasing demand and identify people who may require training or additional support. At the product and model level, they can see where more expensive capabilities are being used and whether that demand persists. These views are intended to support decisions about investment, enablement, coaching, and limits.

The article warns that selecting the model with the cheapest tokens can increase the total cost of a workflow. A less capable model may produce unacceptable results, trigger repeated attempts, or create output that employees must correct. A stronger model may charge more per token but still be more economical if it reaches an acceptable result sooner, requires fewer retries, and reduces human review. Models should therefore be compared by the full cost of producing an accepted outcome, not by their listed token rates.

To make that comparison meaningful, organizations should create evaluations based on real work, including difficult and unusual cases, and define the minimum acceptable quality before testing begins. They should then measure model and tool consumption, the number of attempts, completion rates, latency, and the amount of human review required. For important workflows, the preferred metric is cost per accepted result. In customer support, the result might be a resolved case; in software engineering, it might be a tested change that passes review. That cost should be considered alongside business benefits such as shorter cycle times, protected revenue, avoided risk, saved employee time, or newly available capacity.

Workflow design also affects economics. Clear instructions, narrowly selected tools, reusable context, and explicit stopping conditions can prevent unnecessary loops and reduce wasted consumption. Smaller or faster models are appropriate when they consistently meet the required quality threshold, while the most capable models should be reserved for complex, ambiguous, or high-stakes assignments. This makes model routing a practical management decision rather than a blanket preference for either the cheapest or the most advanced option.

Governance is described as the operational foundation that determines which AI workflows can safely scale. Leaders must specify what company context ChatGPT may use, which connected tools it can access, what actions it is permitted to perform, who must approve higher-risk steps, and how teams can request more capacity after demonstrating value. These questions become especially important when organizations introduce plugins, connectors, Computer Use, and other capabilities that can act across enterprise systems rather than merely generate text.

ChatGPT Work is presented as providing centralized controls over access, approved context, connected tools, permitted actions, usage, and spending. Workspace defaults, group-level limits, individual exceptions, and review requests containing project context can allow administrators to fund promising work without increasing limits for everyone. The proposed approach joins financial control with risk management, so higher spending can be approved selectively when a workflow has a credible business case and suitable safeguards.

For priority deployments, OpenAI says its AI Deployment Engineers can assist customers with evaluations, architecture, latency, reliability, workflow design, performance, and cost efficiency. Privacy and governance are meant to be addressed from the beginning rather than added after deployment. Sensitive workflows require appropriate access controls, retention policies, compliance visibility, and approval paths before expansion. The article also points to enterprise privacy features, including Zero Data Retention options where applicable, for deployments in environments with stringent trust requirements.

Finally, enterprise AI investment should be managed as a portfolio rather than as one uniform program. The article distinguishes broad access for routine employee productivity, specialized workflows that improve repeatable work within particular functions, and a smaller set of strategic initiatives built around proprietary company context. This portfolio framing gives leaders a way to balance widespread experimentation with disciplined investment in workflows that can demonstrate repeatable outcomes, operational readiness, and meaningful business value.