Published on the 22/07/2026 | Written by Heather Wright

A playbook for cost control…

Nobody expects a US$55 trillion cloud invoice. But as agentic AI usage accelerates across business, many organisations are discovering a less spectacular, but equally uncomfortable, reality: AI costs are rising faster than expected.

While model prices are falling and capabilities are improving, businesses are handing increasingly complex work to AI agents. What starts as a few prompts quickly becomes research projects, workflow automation, reporting, coding and decision support. The result is a new challenge for CIOs and CFOs: How do you keep AI costs under control when adoption is exactly what you want?

“The real metric is ‘useful work per dollar’, not simply the cost of individual AI transactions.”

A recent report from cloud finops and spend management company DoIt found 79 percent of the 500 US and UK enterprises surveyed were experiencing AI cost overruns. Meanwhile, Flexera’s 2026 State of ITAM report tells a similar story, saying  just 31 percent of organisations have visibility into their AI software, while 59 percent report increased wasted AI spend.

As iStart reported recently, the AI free lunch is starting to look not so free anymore. Uber revealed it had exhausted its entire 2026 AI coding budget in just four months, driven by token-maxxing where engineers were encouraged to maximise usage. Another unnamed company reportedly accrued a US$500 million bill for Anthropic Claude use in a single month after giving staff unrestricted access and uncapped API token usage for agentic workflows.

That shift has prompted OpenAI to publish guidance on managing AI investments in the agentic era.

Cheaper AI, bigger bills

One of the more interesting observations from OpenAI’s guidance is that leaders should stop obsessing over token prices and start focusing on outcomes. The company argues the real metric is ‘useful work per dollar’, not simply the cost of individual AI transactions.

That matters because lower prices do not automatically translate into lower spending. In fact, history suggests the opposite. When cloud storage became cheaper, businesses stored more data. When cloud computing became more affordable, they ran more workloads. When video meetings became easier, calendars filled and network infrastructure requirements grew.

AI is following the same pattern. As employees gain confidence, they use it for bigger projects, more complex tasks and increasingly business-critical processes. Lower costs remove barriers to experimentation, which often accelerates adoption rather than reducing the bill.

Start with visibility

OpenAI’s first recommendation is visibility. Leaders need to know who is consuming resources, which models are being used, which departments are generating demand and what business outcomes are being achieved.

“Without that visibility, a growing bill is hard to interpret. It could reflect waste, productive experimentation, or a workflow that is starting to become business-critical,” OpenAI says.

Outcomes, not activity

The second recommendation is efficiency. OpenAI warns that cheaper models may not always produce the best result if they require additional attempts, generate lower-quality outputs or create work requiring extensive human review. A more capable model that gets the answer right first time may be cheaper overall than a budget option that needs repeated prompting.

Rather than measuring prompts, queries or tokens, organisations should focus on outcomes such as cost per customer case resolved, contract renewed, report generated or software feature delivered. The goal is to match the model and workflow to the task: use smaller or faster models when they meet the quality bar, and reserve frontier intelligence for complex, ambiguous or high-stakes work.

Governance, growth and getting serious

Governance is the third plank. OpenAI argues organisations need clear policies and controls as AI adoption grows, and that governance should work as “the operating layer that determines which AI work can scale”. That includes expectations around approved tools, data usage, security requirements and oversight of high-impact workflows, particularly as agentic systems take on more autonomous tasks.

OpenAI also recommends investing in AI workflows that become more valuable over time. Rather than spreading spending across dozens of disconnected experiments, it suggests prioritising repeatable processes that can scale across teams, improve with use and create lasting operational benefits.

The final message is simple: do not assume today’s AI usage levels are the ceiling. As agentic AI becomes more capable, employees will find new ways to use it, driving additional demand across the business. Rather than being surprised by rising consumption, organisations should expect it, budget for it and put visibility and controls in place early.

In other words, if the AI bill is growing because people are finding genuine value in the technology, that is not necessarily a problem. The challenge is knowing the difference between productive growth and expensive curiosity.

Which brings us back to that US$55 trillion bill. An AWS billing system fault last week generated estimated invoices ranging from billions to trillions of dollars. The estimates were wildly incorrect, but the broader lesson stands: Technology costs are easy to ignore right up until they become impossible to explain.