OpenAI published a five-step guide this week for enterprise leaders on how to control AI spending as the tools move beyond simple chat interfaces into agents that can run tasks, access company systems, and operate for hours at a time. The guide reads like product documentation for OpenAI's own admin tools, but the problem it addresses is real and not limited to OpenAI customers.
The cost risk here is not theoretical. Uber gave roughly 5,000 engineers access to an AI coding agent in late 2025. By April 2026, the company had burned through its entire annual AI budget, with 84% of developers classified as heavy users before any financial controls were in place. A separate large enterprise reportedly spent $500 million in a single month after deploying AI access with no usage caps. Microsoft reportedly began canceling internal AI coding licenses after facing similar runaway costs.
The reason this happens is structural. Traditional software has a fixed price: you buy seats, assign licenses, and the bill is predictable. AI agents work differently. Every action an agent takes, every document it reads, every step in a workflow, generates a charge. One employee request can trigger dozens of hidden charges that only become visible when the bill arrives. There is no way to manage that without visibility into what is actually running.
OpenAI's new enterprise admin tools, released in June, let managers see credit usage broken down by team, individual, model, and type of work. Employees can see their own spending against a budget and request more capacity when needed, with context about what they are working on. This is a step forward, but analysts at Forrester note the key limitation: these dashboards measure consumption, not outcomes. Seeing that your procurement team is a heavy AI user tells you nothing about whether supplier negotiations have improved or invoice processing times have dropped.
The more useful mental shift is to stop treating AI spending as a cost to minimize and start treating it as a tool to allocate. Some workflows genuinely justify higher spend because they save hours, reduce errors, or protect revenue. Others generate activity with no clear business return. The only way to tell the difference is to define what a good outcome looks like before a workflow goes into production, then measure it consistently.
Pricing has also dropped significantly. OpenAI says the cost per million units of AI output fell 97% between GPT-4 and GPT-5.4, and the newest model family continues that direction. GPT-5.6, released July 9, comes in three versions at very different price points: the most capable version targets complex tasks, a mid-range version covers everyday work at roughly half the cost, and a faster lightweight version handles high-volume simple tasks at the lowest price. Choosing the right version for each workflow, rather than defaulting to the most powerful option everywhere, is itself a meaningful cost decision.
Gartner forecasts worldwide AI spending will reach $2.59 trillion in 2026, up 47% from last year. That number captures the full stack of AI costs across infrastructure and software, but the pattern it reflects is consistent with what individual businesses are experiencing: spending is rising faster than the governance structures needed to justify it. According to PwC, only 12% of CEOs say AI has so far delivered both cost and revenue benefits simultaneously.
The practical question for any business operator is not whether to use these tools. It is whether your team can answer two questions before deploying an AI agent: what does a successful outcome look like, and how will you measure it? Companies that build that discipline early, before agent use scales across departments, will find it far easier to defend their AI budgets to finance and to identify which investments deserve more resources.