AI is supposed to automate accounting. It still can’t get every accounting task right

by
Share this story

KEN GRIFFIN IS RIGHT – AI IS A BUBBLE!

Here is just another example – one of the most widely touted AI use cases is accounting automation.

And it may be one of the clearest examples of where AI still falls short.

Someone finally benchmarked AI on real accounting work.

DualEntry tested leading models across actual workflows across reconciliations, reporting, month-end close, and more.

Every model failed at least a third of the tasks.

Even the best-performing model got roughly 1 in 3 wrong.

A 66% hit rate would be unacceptable in almost any other professional setting.

Yet we’re already being told AI can automate finance teams!

And the most dangerous failures aren’t obvious.

Imagine AI classifies a capital expenditure as an operating expense.

The entry looks perfectly normal.

No warning or obviously error just confidently wrong.
But now your profitability is wrong, your tax position is wrong and every report built on that entry inherits the mistake.

You might not discover it next week, you might discover it three months later when something doesn’t tie – and now someone has to go back through everything.

Other research this year has suggested that a significant portion of the time saved using AI in finance gets spent correcting its output!

That doesn’t surprise me.

I’m not anti-AI, I use these tools constantly.

They’re fantastic at drafting, working quickly, finding information and surfacing things that might otherwise take hours.

But the job hasn’t disappeared, it’s just freed up some time e.g. less time entering the data and more time figuring out whether what AI produced is actually correct.

Maybe the real skill is spotting a confidently wrong answer, case in point below:

“You’re absolutely right, and I’m sorry.”

Not exactly what you want to see halfway through a month-end close.

0 views