The AI ROI Problem: Why Firms Can't Prove AI Spend Works

Introduction

National Treasure (2004) may not have been a critical hit, but it sure was engrossing and entertaining. It had Ben Gates (Nicolas Cage) unraveling clues that led to historical artefacts and then even more ancient artefacts.

It seemed like an endless loop of finding more clues with no ultimate payoff. Instead, Gates was just burning through cash tracking down clue after clue, becoming problematic even with the backing of wealthy investors. That was until he finally found the treasure.

The story is the same for artificial intelligence (AI) companies at this point in the AI race. They’re burning through cash to build data centers, buy GPUs, and boost their infrastructure to expand their capabilities and stay ahead in a cutthroat industry.

The problem is they’re spending faster than they can earn, and with the availability of free or cheaper tools, pricing models at premium prices come with the risk of losing customers. At the same time, they must keep spending to maintain a leading position in the AI market.

This is a classic ROI Problem. Let’s explore whether the treasure is missing, hidden, or simply taking longer to uncover.
 

What Is The AI ROI Problem?

The enterprise AI ROI problem is the gap between what organizations spend on AI and the business value they can prove.

Return on investment usually compares a project’s financial gain with its total cost. That becomes harder when the investment changes many workflows, and its gains appear as minutes saved, better decisions, fewer errors, or risks avoided.

Domino Data Lab’s 2026 survey of 639 senior enterprise AI leaders captures the contradiction. While 93% reported improved AI production capabilities, 57% said returns still failed to outpace investment. Frends found a similar disconnect: only 25.5% of surveyed AI pilots or deployments had produced measurable profit-and-loss impact.

The technology may be producing something useful, but the organization cannot always connect that output to financial records. Before declaring the treasure missing, it helps to see where the money went.
 

Why Is AI So Expensive To Build And Scale?

Building AI and GenAI (generative artificial intelligence) requires specialized chips, data centers, energy, technical talent, model development, security, and continuous testing. Furthermore, costs continue after launch because every model call, retry, and agent action consumes computing resources.

Enterprises inherit another layer of expenses. Their AI expenditure includes subscriptions, cloud services, data preparation, integrations, governance, training, workflow redesign, monitoring, and human review.

Ahead of this, the hidden cost of data centers leaves us with another question that needs to be answered: Is AI Draining Our Water Supply?

CloudZero notes that AI costs move across providers, products, customers, features, agents, and workflows, while the final bill may show what was spent without showing what it produced.

Older systems make scaling harder. Frends reported that 59% of surveyed organizations struggled with point-to-point integration complexity, while 65% struggled to connect legacy and cloud platforms. The model may be ready, but the business around it often is not.

That leads directly to the measurement problem.
 

Why Are Enterprises Struggling To Prove AI ROI?

Companies often track AI activity instead of business outcomes.

Token use, prompt volumes, chatbot sessions, generated documents, and hours saved can show that people are using GenAI, but they do not prove higher revenue, lower costs, better retention, or stronger margins.

Pendo argues that token volume confirms activity, not whether employees used AI for valuable or underdelivering work. SVPG describes a similar productivity paradox: teams may produce more and move faster without improving the result customers or the company need.
 

Why Enterprises Fail To Measure Generative AI Returns

Many projects begin without a baseline, an accountable owner, or a specific business metric. Broad copilots are easy to distribute but difficult to evaluate across hundreds of small tasks. Weak data, disconnected systems, poor adoption, and pilots that never enter daily operations widen the gap.

Lanai’s 2026 AI Labor Report found that 90% of surveyed organizations lacked a dedicated function for tracking ROI, while 88% had no formal method for attributing business outcomes to AI. When ownership is scattered, the evidence is scattered too.
 

How Financial Officers Evaluate Generative AI Business Value

Financial officers look for changes in revenue, operating cost, margin, risk, customer behavior, or labor capacity. That means measuring return on AI investment at the workflow level and comparing performance before and after adoption.

However, AI-assisted work may be credited entirely to an employee or buried inside a larger process. Here, 87% of respondents sometimes or always credited AI-assisted output entirely to people. The value exists, but its origin disappears.

That raises another possibility: the return may be real but hidden.
 

What AI Benefits Are Traditional ROI Metrics Missing?

Traditional ROI favors benefits that can quickly be expressed in money. AI often creates indirect gains first, including faster research, shorter response times, improved knowledge access, fewer routine errors, and more capacity for complex work.

Those improvements do not become savings automatically. Saving ten hours creates no financial value if the time is absorbed by meetings, revisions, or more low-value output. It matters when the recovered capacity improves a measurable result.

Shadow tools can also help employees draft, analyze, classify, or summarize without appearing in approved technology records. This is known as “AI labor orphaning,” where machine-assisted output never enters the systems used to assess performance or financial results.

Therefore, proving enterprise AI value requires tracking direct gains and what employees do with the speed, time, or information AI provides.

Still, even strong measurements cannot make every return appear immediately.
 

Why Can AI ROI Take Years To Appear?

The first stage of an AI program often builds capability rather than profit.

Companies clean data, connect systems, establish governance, train employees, test use cases, and redesign workflows before larger returns arrive. Early gains can compound when organizations reinvest recovered capacity into automation.

This makes generative AI ROI appear weak when leaders apply a short payback window to a long organizational change. Some pilots fail and should be stopped. Others create infrastructure that later projects can reuse at lower cost.

Like Ben Gates following clues, enterprises may spend heavily before reaching the treasure. The answer is not endless patience. It is separating a promising foundation from an expensive dead end.
 

How Do Companies Calculate ROI On Generative AI?

Companies should begin with one business problem, a baseline, and a defined owner.

They must record the full cost, including technology, cloud usage, integration, data work, security, training, supervision, and maintenance. They can then compare it with revenue, reduced processing costs, fewer defects, faster cycles, lower risk, or stronger retention.

For measuring return on AI investment, the basic formula remains: financial benefit minus total cost, divided by total cost, multiplied by 100.

Leaders should also track adoption, quality, reliability, and future capability. These indicators show whether an early investment is moving toward financial value or merely producing activity.

The strongest approach ties each use case to the workflow it changes, reviews results over several time horizons, and redirects funding when evidence remains weak. That is how companies calculate ROI on generative AI without pretending every benefit is immediate or impossible to quantify.

The final clue is knowing which investments deserve more.
   

Conclusion

The AI ROI problem is not one problem. Some investments are weak, some returns are hidden, and others need time to mature. Enterprises should neither accept endless spending on faith nor cancel foundations too early.

Like the treasure hunt in National Treasure, the clues matter only when they lead somewhere, even if just to the next promising clue. The winners will be the companies that connect every AI cost to a measurable business outcome.