This article is a guest post by the Chief Financial Officer of DualEntry.Woosung Chun. The views expressed in this article are solely those of the author.

We are building AI-powered accounting software. That's precisely why we need to figure out where AI's shortcomings in accounting truly lie.

We tested 19 mainstream AI models on 101 real accounting workflows. The tests were not common-sense questions or multiple-choice questions like "What is accounts payable?" but real accounting scenarios: classifying this transaction, creating a journal entry for this scenario, reconciling this bank statement, completing this month-end close—these are exactly the types of tasks every finance team encounters in daily work.

The best-performing model in the tests achieved an accuracy of 79.2%, which was Claude Opus 4.7. In second place was OpenAI GPT-5.4 with an accuracy of 77.3%. For comparison, GPT-4 scored only 39.8% on the same set of tasks. Regardless of your views on AI, this trajectory of progress cannot be ignored.

But one number keeps lingering in my mind: the world's most powerful model still gets accounting tasks wrong one out of five times.

In most scenarios, 80% is a decent score. But in accounting, it's far from sufficient. A misclassified transaction doesn't just stay in place; it flows into the income statement, balance sheet, tax filings, and even documents handed to auditors. An error during month-end close doesn't just affect one account; it compounds and amplifies.

What surprised me most was not the overall accuracy, but where the models failed. Models that performed reasonably well on conceptual accounting knowledge questions (like "Do you understand accrual accounting?") often collapsed on bank reconciliation and month-end close tasks. The most difficult operational work—the content that truly keeps finance teams up at night—is precisely where AI's performance drops most noticeably.

There is a fundamental difference between "knowing accounting" and "doing accounting." I was aware of this before the tests, but I didn't expect the gap in the data to be so stark.

Another point worth noting: most AI tools currently being pitched to CFOs have not undergone such evaluations. Vendors showcase demos and outputs that look correct, but they don't show task-level failure rates on a representative sample of real workflows. There is currently no unified standard in the industry, which is why we built and publicly released the2026 Accounting AI Benchmark—the full methodology and result set are openly accessible.

So what does this mean in practical terms? There are a few points worth noting.

First, any CFO considering AI in finance operations should require vendors to provide task-level accuracy for their specific workflow categories, rather than overall benchmark scores. Our data shows that differences between categories are quite significant. A model that performs reasonably well on transaction classification may fail badly on bank reconciliation. The risk profiles of these two are completely different.

Second, the mechanisms for handling model errors are just as important as how often the model is correct. A 20% error rate is not unmanageable—provided you have validation layers, review processes, and controls between AI output and the books. Without these, a 20% error rate becomes untenable. But the answer is not to avoid AI in accounting; rather, it is to carefully design where the model sits in the workflow. There is a fundamental difference between AI drafting and suggesting, embedded in a system with core functions like deterministic validation, audit trails, and exception handling, versus simply bolting AI onto legacy systems or calling it raw via API without any infrastructure. The former is viable; the latter allows errors to silently accumulate until they erupt all at once.

Third, any financial AI tool that cannot tell you its error rate is suspect. Not because vendors are dishonest, but because most vendors indeed have not tested in this manner. Task-oriented evaluation against real accounting workflows is much harder than building product demos. This is not a criticism—task-oriented evaluation is genuinely difficult to construct.

But the gap right now is real. The progress from GPT-4 to today's top models over two years is remarkable. I don't believe 79.2% accuracy is the ceiling; it's closer to the lower bound of where frontier models will be in 12 months.

However, model capability and deployment readiness are two different things. The controls, validation processes, and audit trails needed to run AI safely in real accounting environments take time to build and test. For now, model capability is outpacing the systems designed to catch its errors.

This gap is exactly what CFOs need to watch closely.

The benchmark was designed and built byIgnacio Brasca, a senior software engineer at DualEntry.