OpenAI Says Agent Budgets Should Be Managed by Process, Not Department

OpenAI published five steps for managing investment in agent systems. The recommendations reflect a change in operations: a normal conversational tool produces a few responses, while an agent may search for hours, call other programs, process files and repeat failed steps. Costs can therefore rise without any direct link to employee numbers. The first step is detailed visibility into usage and spending. Consumption should be divided by model, workflow, team and task. A single monthly invoice does not show whether spending came from a useful process or from an agent repeatedly returning to the same error. The second step is to assess efficiency by outcome. The cheapest model may not produce the lowest total cost if it needs more attempts or supervision. A more expensive model may complete the task on the first run. Financial comparison must therefore include completion, quality and employee time. The third recommendation is to govern advanced workflows before scaling them. An agent that only reads public websites presents a different risk from one that sends emails, places orders or changes company databases. Organisations must define permitted tools, financial limits and steps that require human approval. The fourth step is to fund processes with cumulative value. A one-off presentation may save several hours. An agent that continuously reviews contracts, updates customer records or detects code defects can create value repeatedly, but it requires more stable integration and regular review. The fifth step is to match capacity to proven demand. OpenAI warns against buying large amounts of computing power too early. Companies should first prove that a workflow delivers results and only then negotiate dedicated capacity or long-term commitments. “Improve visibility into usage and spending,” is the first recommendation. Another says to “govern advanced workflows before scaling.” Both formulations show that an agent system can no longer be managed like ordinary office software. Chief financial officers face a new cost-allocation problem. One agent may work simultaneously for sales, legal and customer service. The invoice therefore cannot be divided simply by the number of user accounts. It makes more sense to assign costs to the process and its business owner. Agent governance also has a legal dimension. European AI rules base some obligations on the purpose of use, not only on the underlying model. The same technical system may pose low risk when summarising internal material and higher risk when evaluating employees or deciding access to a service. Companies must therefore record where and why an agent has been deployed. A central process catalogue should include the owner, data used, external tools, possible harm, human oversight and the date of the last review. Without such a register, costs cannot be managed and accountability cannot be demonstrated.

OpenAI Says Agent Budgets Should Be Managed by Process, Not Department

The full article is available to users registered at Hard Skills