Cloud Asked What It Cost, AI Is Asking What It's Worth | Harness Blog

A year ago, most of my customer conversations were about cloud cost attribution, commitment coverage, and rightsizing. Classic FinOps stuff. The occasional surprise bill, sure, but nothing that kept anyone up at night. I stopped asking “what was that ‘oh $*#@’ moment when you saw a bill and knew something had gone seriously wrong?” because no one had one of those moments anymore.

Somewhere in the last 12 months, the tables turned and our customers were asking me a serious question: why did my AI bill just do that, and how do I stop it from happening again?

Clearly, something changed. So we ran a study to figure out why.

We surveyed 700 engineering leaders and practitioners across five countries this spring to ask about their organization's FinOps practices. All respondents are at organizations with at least 1,000 employees and real, recurring AI spend, not startups experimenting on a credit card. What we found in the 2026 State of AI in FinOps survey report reads less like a new problem and more like an old one wearing a different jacket.

The pattern I keep seeing

I spent years in FinOps for cloud before AI spend was a line item anyone thought about. But the patterns in the AI billing data – the ownership confusion, the governance gaps, the invoice showing up before anyone understood why – are the same ones we saw in cloud 10-plus years ago. Now, however, the patterns are compressed into a much shorter timeframe and coming at an accelerated pace, if you can believe it.

67% of organizations now spend more than $250,000 a month on AI. 20% have already crossed $1 million a month. At that scale, AI isn't a tool cost anymore, it's a capital decision, which deserves the same governance and attribution discipline it took the industry a decade to build for cloud. We don't have a decade this time.

Ownership is the root, not the symptom

If you ask five people in a typical enterprise who owns AI costs, you'll get five different answers. 52% of respondents told us there's no clear, dedicated AI cost owner in their organization. And the deeper issue isn't that nobody's watching, it's that four different functions are each contributing to the bill. Platform and DevOps teams carry 30% of the accountability, with FinOps 27%, finance 23%, and engineering 19%. No single function comes close to a majority and that is one of the top problems enterprises are facing at the very start.

Meanwhile, engineering and platform teams hold the most influence over spending decisions, more than any function's share of accountability. The disconnect between decision-making and accountability is where a single cost overrun escalates into a full-blown P&L problem. 

When the bill doubles overnight

72% of organizations have hit an unexpected AI cost spike or surprise bill in the past year. 1 in 3 have been caught off guard more than once. That alone wouldn't worry me as much if diagnosis were fast. It isn't. 79% report needing a full day or longer to trace a spike back to its source; and roughly a third need a full week. Meanwhile, the bill keeps running.

In the cloud, knowing a spike happened is only the first indication of a problem. Understanding the cause in real-time is where most organizations still fall short, and in today's AI-driven world, the stakes are higher because the spend curve is steeper.

The waste number that should get a CFO's attention

Across the dataset, organizations estimate that 26% of all AI spend is wasted — consistently, across infrastructure, software, models, and services alike. For a company spending $1 million a month, that's $260,000 a month with no measurable return. Just imagine what any enterprise could achieve with that amount of wasted money turned into investments?

Unlike waste in a traditional cloud landscape (are we really, already, calling cloud “traditional” or “legacy”?), this isn't an idle instance somebody forgot to shut down. It's embedded in usage patterns — prompt design, model selection, retry logic — none of which currently carries a cost signal. 56% of engineering leaders told us their teams don't have cost in mind when building AI features, and only 45% of engineers, on average, understand what their builds actually cost.

That's not a motivation problem. In our FinOps in Focus Report 2025, 62% of developers said they wanted more control over cloud costs, and there's no reason to think AI is any different. The problem is a lack of visibility, not a misalignment of intent. The data simply isn't in front of people when the decisions get made.

Policy on paper, not in practice

The instincts are right; the operational muscle isn't there yet. 73% of organizations say they have an AI cost policy. Only 47% fully enforce it. That's a 26-point gap between what's written down and what actually happens day to day. Policies that don't show up in the tools engineers actually use tend to get ignored, because the incentive to move fast usually wins.

Only 21% describe their AI cost management as fully mature company-wide, and only 26% have a robust way of measuring the business value of their AI spend at all. Four in five organizations simply aren't there yet.

‍Building an AI ROI Culture

What I keep coming back to is that the hardest part of this isn't technical. It's cultural and organizational. The tooling is catching up. The harder work is getting engineering teams to build with cost in mind from the start, not as a constraint, but as a design principle.

The organizations in our data that did reach maturity didn't try to fix everything at once. They followed a fairly consistent sequence: 

• name an owner before buying a tool, 
• build a unified cost view before optimizing, 
• embed cost data into engineering workflows before writing policy, 
• and establish unit economics tied to real business outcomes before chasing ROI.

Use this report to find out where your organization stacks up against others in the industry and how mature enterprises are getting ahead of AI cost blindspots..

Download and learn more about the 2026 State of AI in FinOps.