The AI race shows no sign of slowing down. The four hyperscalers are on track to spend more than $725bn on AI infrastructure in 2026. That buildout is being paid for somewhere, and as we’re all seeing, a growing share of it shows up downstream in our enterprise AI bills.

The problem is that while it’s easy for organizations to see how much they’re spending on AI – it’s far more difficult to track what they’re getting in return.

Harness' new 2026 State of AI in FinOps report, based on a survey of 700 engineering leaders and practitioners across five countries, shone a spotlight on the challenges we’re all facing in this area.

It found that AI spend is moving faster than organizations' ability to track, attribute, or govern it. The result is a growing blind spot. A quarter of enterprise AI spend is estimated to be wasted, while unexpected cost increases are leaving teams struggling to understand what happened.

AI costs are rising faster than organizations can explain

Two in three organizations now spend more than $250,000 a month on AI, and 80% say that investment has increased over the past six months. AI has moved from an unexpected expense to a recurring business cost that organizations need to actively manage.

But understanding where that money is going is becoming increasingly difficult.

AI costs are spread across infrastructure, software, models and managed services, often between multiple providers with different pricing models. A single organization may have model usage, developer tools, and enterprise AI platforms all contributing to the same bill.

Meanwhile, the biggest costs don’t always show up in the places organizations expect. Productivity tools such as coding assistants and enterprise AI platforms account for a significant share of spend, but often look more like software licenses than infrastructure costs. As a result, they often sit outside the processes organizations built to manage cloud spending.

The result is a growing visibility gap. When costs increase, organizations can see the impact on their bills but have little insight into what actually caused the jump.

Nobody owns the bill

AI costs remain difficult to explain because ownership is still unclear. More than half of organizations (52%) say there is no dedicated owner for AI costs across the business. Responsibility is spread across engineering, finance, and IT – meaning multiple teams add line items to it, but nobody is accountable for the sum total of the bill.

This differs from previous technology investments. Cloud costs are closely tied to infrastructure – servers, storage, and networking – making ownership easier to assign. AI costs work differently. They are shaped by engineering decisions, product choices, model selection, and user behavior all at once. The person making the decision that increases costs is not always responsible for explaining them later.

That gap creates a costly delay when something goes wrong. Nearly three-quarters (72%) of organizations have experienced an unexpected AI cost increase in the past year, but only 20% say they could identify the cause within hours. For many teams, answering “why did the bill spike?” can take days.

Tokenmaxxing – when more AI does not equal more value

The cost management problem is not just about visibility and ownership. It’s also about the incentives organizations create around AI usage.

During early adoption, encouraging teams to use AI as much as possible helped employees experiment and build confidence. But as AI becomes a significant business cost, maximizing usage becomes a much less useful measure of success.

Despite this, 57%, of respondents say their organization encourages “tokenmaxxing” – pushing employees to maximize AI usage regardless of the value it creates.

The problem is that many engineers still don't know what the AI features they build actually cost to run. Less than half (45%) say they understand those costs, meaning decisions around models, prompts and workflows are often made without considering the financial impact.

When those costs are hidden, developers can choose more expensive options, run more experiments and generate more output without seeing the consequences. It’s unsurprising that an estimated 26% of AI spend delivers no measurable return.

The issue is that engineers are being asked to make spending decisions without the information needed to ensure they make good business sense.

What needs to change

As we all know, none of these challenges are new. The cloud industry spent years solving problems around ownership, attribution, and governance. AI is creating many of the same issues, just on a much shorter timeline.

The starting point to addressing the issue is assigning clear ownership for AI spend. Giving one team ownership creates a clear starting point for every other improvement.

From there, organizations need to create a single view of AI spending across providers, models, software, and infrastructure. It’s difficult to manage costs when the data arrives in fragments from multiple systems.

Finally, cost data needs to become part of the engineering workflow, not something finance reviews after the invoice arrives. Engineers make decisions about prompts, models, and deployments every day, yet those decisions are often made without any visibility into the cost implications.

The organizations making the most progress are looking beyond spend alone. They’re connecting AI costs to business outcomes, establishing clear unit economics and measuring where AI creates value rather than simply how much it costs.

See it or lose it

For most organizations, the decision to invest in AI has already been made. The challenge now is making sure that the investment is delivering what the business needs. That starts with understanding where the money is going, who is responsible for it, and whether teams have the information they need to make wise decisions.

What measures does your organization use to trace its AI spend today – and would everyone agree on who owns it?