AI’s ROI reckoning: from blank checks to business results.

Across boardrooms and budget meetings, AI is now a line item that must justify itself. Finance chiefs who once championed ambitious AI bets are confronting hard facts: 87% of senior finance executives say they need to tie AI investments to concrete business outcomes within the next year, yet only 22% can currently do so. This “ROI gap” leaves many CFOs in a bind—already 60% of finance leaders admit they’re spending more on AI than they can rationalize. Boards have noticed. Only 26% of corporate boards still support a “blank check” approach to AI, while the other two-thirds now insist on evidence of return before approving new projects. In fact, 22% of boards won’t authorize any additional AI spend until they see proven ROI in black and white.

CFOs are responding by raising the bar for AI business cases. No longer is it sufficient to cite nebulous “productivity improvements” or promise long-term disruption. They want to see credible forecasts of cost savings, revenue growth, or efficiency gains—and plans for how those will be measured. Finance teams are rolling out new frameworks to monitor AI expenses and benefits more rigorously. For example, finance FinOps firm CloudZero found that 75% of organizations unable to quantify AI outcomes have already halted further investment in some initiatives, and 35% have killed projects outright due to lack of measurable impact. By contrast, among companies that can demonstrate ROI, only 11% have had to ax an AI project. The implication is stark: if you can’t prove value, your AI project may not survive the next budget cycle.

This scrutiny is creating tension between IT and finance teams. In 74% of enterprises, AI initiatives are still driven outside of the finance department’s control—but it’s the finance team that gets stuck paying the bills 60% of the time. CFOs find themselves accountable for costs they didn’t directly approve. That dynamic is accelerating the push for ‘cost-to-value’ transparency. Forward-looking organizations are establishing cross-functional “AI value” committees and requiring that every new AI pilot define success metrics upfront, whether it’s reduced fraud losses, increased customer acquisition, or faster product delivery. The goal: to reconnect the excitement of AI with the fundamentals of business value.

Even as internal scrutiny intensifies, external market forces are shifting AI economics in enterprises’ favor—and raising new strategic questions. In the past 48 hours, the two leading AI model providers, OpenAI and Anthropic, engaged in a rapid-fire price war, each slashing their flagship model fees by 40–50%. This unheard-of dual price drop, coming on the heels of cheaper “mid-tier” AI systems entering the market, is a clear sign of mounting competitive and customer pressure. For large AI users, these cuts offer immediate budget relief. However, they also underscore how quickly yesterday’s expensive cutting-edge service can become today’s commodity. Vendors are racing to retain customers like that $15 billion legal tech startup that famously saw its gross margins turn negative 50% due to hefty OpenAI and Anthropic API costs. That startup responded by developing its own AI model (nicknamed “Kimi”) to reduce dependency on external providers—and in doing so, restored its margins virtually overnight.

The episode highlights a broader trend: companies are rethinking the “buy vs. build” calculus for AI. While outsourcing AI capabilities to cloud vendors allowed for speedy initial wins, the bill can quickly become unsustainable at scale if usage isn’t controlled. Unlike fixed software licenses, most AI services charge by consumption—every API call or million tokens adds to the tab. If employees enthusiastically adopt a new AI tool across the organization, costs can skyrocket ahead of any commensurate uptick in revenue or savings. This “pay-as-you-go” model puts AI-intensive startups and business units in a catch-22: pull back on usage and you stunt AI’s benefits, push forward and you might burn through your budget with little to show for it.

To escape that trap, more firms are exploring hybrid approaches. Some are negotiating better rates and committing to one primary AI platform, while others are taking a diversified route. The latter approach involves “model routing” strategies: dynamically allocating AI workloads to whichever provider offers the best performance-per-dollar for a given task. It requires sophisticated MLOps engineering, but it can tame costs and avoid lock-in. And as the $500 million billing scare demonstrated, strong governance is essential regardless of vendor. Setting strict usage limits, real-time monitoring of token consumption, and automating cost controls are becoming standard requirements before scaling any AI solution. In short, enterprises are learning that controlling AI costs is as critical to ROI as boosting AI capabilities.

Amid the sobering news, there are bright spots where AI investments are yielding tangible returns. At a Reuters AI summit, FedEx revealed it has achieved more than $3 billion in cost reductions over six years by deploying 200+ AI and data use cases throughout its operations. Consumer goods giant Mars shared that AI-driven insights led to a roughly 20% increase in sales through Amazon’s digital channel. And job platform Indeed reported that intelligent matching now drives about 70% of all hires on its site. These successes aren’t luck—they come from targeting AI at clearly defined problems (like logistics optimization, e-commerce conversion, and matching job seekers to openings) where improvements can be directly measured in dollars or key performance metrics.

What do the winners have in common? First, they pick their battles wisely. High-ROI AI projects tend to focus on domains with quantifiable outcomes: fraud reduction, supply chain efficiency, customer service resolution rates, software productivity, etc.. These are fields where it’s easier to isolate the impact of AI—for example, if an algorithm catches fraud that would have been missed, or shaves 20% off delivery times, those savings or gains can be calculated. Second, successful organizations treat AI projects like business transformation initiatives, not science experiments. That means redesigning workflows and roles to capture the full benefit of AI assistance, rather than simply layering new tools onto old processes. As one AI strategist put it, “Until we redesign our architecture around AI, we’re not going to see the full value”.

Crucially, leaders who achieve ROI enforce a culture of measurement and accountability. New metrics are emerging to track AI value more rigorously. For instance, KPMG’s head of AI labs advocates measuring “cost per accepted output”—the total cost to produce one useful result with AI, including model fees, human review, error handling, and fallbacks. This kind of end-to-end cost metric helps teams identify when the overhead of using AI outweighs the benefits. If an AI system requires so much human correction that it erodes efficiency, the problem is transparent.

Finally, high-performing organizations are unafraid to pull the plug on underperforming initiatives. At the AI summit, speakers from industrial manufacturer AGCO and genealogy firm Ancestry stressed the importance of terminating AI projects that fail to deliver value, rather than letting pet projects linger out of habit. This is a marked shift from earlier years when any AI pilot—no matter how dubious—might have been kept alive for learning’s sake or optics. Now, if an AI project can’t demonstrate real impact on the business, it may be better to walk away and reallocate resources to more promising areas. Leaders are learning that saying “no” or “not yet” to certain AI proposals is sometimes the smartest way to ensure the investments they do make will produce meaningful returns.