Why Cost Accounting is the Wrong Way to Value Artificial Intelligence
More than 80% of companies report zero measurable impact from AI on employment or productivity over the past three years.
Zero.
That’s from a National Bureau of Economic Research study of nearly 6,000 executives across four countries. And yet those same executives expect AI to boost productivity by 1.4% over the next three years.
Economists are calling it Solow’s paradox: the same gap between investment and results that defined the early computer age. You could see computers everywhere but in the productivity statistics.
The standard advice is to be patient. Transformative technologies take time to show up in the data. I get it. But I also agree with Alexander D. Hilton, the author of this CFO.com opinion piece: most firms are thinking about AI the wrong way.
The Cost Accounting Trap
The typical approach to AI focuses on the time a task takes today. Multiply it by the loaded labor cost. Calculate how much faster AI can do it. That factors into the return on investment.
That’s cost accounting logic. And, as Hilton points out, it’s the same logic Eliyahu Goldratt dismantled decades ago with his Theory of Constraints. Manufacturers ran every machine at full utilization because the local math said to.
The result was excess inventory, longer lead times, and hidden costs that never showed up in the numbers. The machines were busy. The system was slow.
The same mistake is playing out with AI right now.
When Klarna replaced 700 customer service agents with a chatbot in 2024, the cost accounting looked great. Until customer satisfaction collapsed and they started rehiring humans.
Making something faster isn’t the same as making the system produce more.
The Right Question: Where’s the Bottleneck?
Goldratt put it plainly: technology only brings benefits if it removes a limitation. The question isn’t whether AI is capable. It’s whether it removes the specific constraint that’s limiting your firm’s output.
That’s throughput accounting. Instead of asking “how much did we save?”, you ask, “how much more can we generate?”
Throughput is the rate at which the organization produces revenue. If an AI initiative speeds up a step that isn’t the bottleneck, throughput doesn’t move.
The CFO.com piece cites a real example: an AI tool accelerated document generation by a factor of 240. Cost accounting would call that transformational. But the bottleneck in the system was subject-matter expert validation. That’s a manual, human-dependent review step that couldn’t be parallelized.
The AI made the non-bottleneck faster. Nothing else changed.
A Firm-Level Reality Check
Right now, AI vendors are pitching tools that can prepare tax returns.
Perplexity just launched a product that drafts full federal tax returns from uploaded documents. TaxGPT says it automated 1040 prep. Basis raised serious money to build AI agents for accountants.
On Episode 482 of The Accounting Podcast, David Leary and I talked through what all this actually means for practitioners.
If your ROI logic is, “AI prepares a return 40% faster, so I save X hours, so the ROI is Y,” you’re probably measuring the wrong thing.
Is tax prep the bottleneck in your firm? In my experience, it isn’t.
The constraints are getting the right documents from clients and the time required for partner review. If AI speeds up data entry and form preparation, but you still have to chase down client documents, and the partner review queue still moves at the same pace, what’s changed? Your junior employees are less busy. But your throughput is identical.
Before you commit to any AI tool, push for honest answers to three questions:
What specific bottleneck does this address? If no one can name the constraint limiting firm revenue, you’re probably optimizing a non-bottleneck.
Does this increase throughput, or just reduce local cost? A process that runs faster but feeds into the same downstream bottleneck hasn’t changed your output.
What new costs does this create? Retraining, integration, validation workflows, or human oversight? If the net effect on throughput after subtracting that overhead is negative, the tool destroys value regardless of how good the demo looked.
Are You Doing the Right Math?
AI can be genuinely transformational for an accounting firm when it’s applied to the actual constraint. I’m seeing it happen.
But the firms that come out ahead will ask the harder question first: where is our constraint, and does this tool actually remove it?
The math for AI works out. You just have to make sure you’re using the right accounting method.
If you want to hear more of our discussion on this topic, check out Episode 482 of The Accounting Podcast.
