This article is a guest post by Alexander D. Hilton (MBA, digital strategy and AI consultant for the Agile New England ACMs chapter). The views expressed are solely those of the author.

A study released in February by the U.S. National Bureau of Economic Research (NBER) surveyed nearly 6,000 CFOs, CEOs, and executives in the United States, the United Kingdom, Germany, and Australia. The results showed that over 80% of companies reported that artificial intelligence had no measurable impact on employment or productivity over the past three years. Yet these executives predicted that AI would boost productivity by 1.4% and output by 0.8% over the next three years. Economists have begun comparing this phenomenon to Solow's paradox—the disconnect between investment and results, as experienced in the early computer age.

The common explanation for this is "patience": transformative technologies take time. This is certainly true, but it is not enough. The deeper problem is that most companies evaluate AI investments using the wrong accounting logic—and unless CFOs correct this, the numbers will not improve.

The Trap of Cost Accounting

Most AI business cases are built on cost accounting: hours saved, hires avoided, cycle times shortened. This logic treats every local efficiency gain as a contribution to profit. It is the same logic that for decades drove manufacturers to run every machine at full capacity—and Eliyahu Goldratt's Theory of Constraints long ago proved that this approach actually leads to excess inventory, longer lead times, and hidden costs.

Today, the same mistake is being repeated in AI. Forrester predicts that companies will postpone 25% of their planned AI spending until 2027, noting that fewer than one-third of decision-makers can link AI value to financial growth. PwC's 2026 Global CEO Survey found that 56% of CEOs said their AI investments have yielded no returns. In 2024, Klarna replaced 700 customer service agents with an AI chatbot, and cost accounting treated it as a victory—until customer satisfaction collapsed and the company quietly began rehiring humans. Forrester also found that 55% of employers who laid off workers due to AI now regret it.

Worse still, some AI deployments not only fail to create value but actively destroy it. Air Canada deployed a chatbot to reduce customer service costs, but the bot fabricated a bereavement fare refund policy, and the tribunal ruled that the airline had to pay compensation. Deloitte Australia used AI to accelerate government reports, yet the deliverables contained fabricated citations and false references, ultimately forcing the firm to partially refund US$290,000 of the contract. In each of these cases, the cost-accounting business case showed lower operating expenses, but failed to capture the increases in investment and operating expenses from legal risks, remediation, and reputational damage—costs that far exceeded the initial savings.

In every case, the AI initiative reduced local costs or accelerated local tasks, but none of them increased the rate at which the organization creates value. This distinction is precisely what Throughput Accounting aims to capture.

A Better Lens: Throughput Accounting

As Goldratt wrote in Beyond the Goal: "Technology can provide benefit if and only if it diminishes a limitation." From this perspective, the CFO's question is not whether AI is powerful, but whether it diminishes the specific constraint limiting financial performance. If it does not, the investment will produce activity, not value.

Throughput Accounting operationalizes this insight through three metrics: Throughput (T) is the rate at which the organization generates money through sales—that is, revenue minus truly variable costs; Investment (I) is the capital tied up in the system, including not only AI licenses and computing resources but also retraining costs, verification infrastructure, and compliance overhead; Operating Expense (OE) is the ongoing cost of maintaining the system, including human oversight, error correction, and governance.

The decision rule is simple: a good investment should increase T, or reduce I and OE—but only if it acts on the system's constraint. If an AI initiative accelerates a non-bottleneck step, system output will not change, the bottleneck does not move, and impressive speed gains will never show up on the income statement.

I have witnessed this firsthand. In a large enterprise transformation project, an AI tool increased document generation speed by 240 times. Cost accounting would treat this as a transformative gain. But the system constraint was expert review—a manual, non-parallelizable step. The result: despite a 240-fold speedup upstream, throughput reached only 89% of the target. AI made the non-bottleneck faster, but the constraint did not move, and the system barely changed.

The Questions CFOs Should Ask Instead

The NBER study reveals an intriguing detail: executives who personally use AI spend only 1.5 hours per week on it, yet approve millions of dollars in enterprise-wide deployments. The gap between executive experience and investment scale suggests that decisions are driven more by vendor narratives and peer pressure than by constraint analysis.

Before approving the next AI project, CFOs should demand answers to three questions. First, what system constraint does the initiative target? If the team cannot point to a specific bottleneck limiting throughput, then the project is optimizing a non-constraint. Second, does it increase T, or does it merely reduce local OE? A process running faster, but still feeding into the same downstream bottleneck, does not increase the rate at which the organization generates money—it just cuts local costs while throughput stays flat. Third, what new I and OE does it create? Every AI deployment carries hidden investments—retraining, integration, verification processes—and ongoing OE such as human oversight and error correction. If the net increase in T minus the increases in I and OE is negative, then no matter how impressive the demo, the initiative is destroying value.

When the Hype Fades, the Math Remains

This is not an anti-AI argument. AI applied to the constraint—truly diminishing the bottleneck that limits throughput—can be transformative. But if applied indiscriminately, evaluated by local speed rather than system throughput, and approved without constraint analysis, it becomes the most expensive way to accomplish nothing. Companies reporting zero productivity impact are not failing because AI is ineffective, but because they are measuring the wrong things. Throughput Accounting provides CFOs with a discipline to tell the difference—and to ensure that AI investments land where they can truly drive performance.