Why Brilliant AI Still Fails at Basic Accounting
The barrier is not intelligence but steadiness, and background software needs to be right almost every time.

The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch · 20VC: Jev Creator on Why Data Is More Important Than Compute | Will US Open-Source Models Catch Chinese | The Cost of Intelligence and Tokens: What No One Sees But Needs to Know with Diogo AlmeidaModern AI can wrestle with math that would stump most people, yet it cannot be trusted to do a company's bookkeeping.
The problem is not smarts, says Diogo Almeida, co-founder of the AI company Jev. It is consistency, or what he calls reliability of intelligence.
He does not mean uptime, the familiar idea that software stays online and responds quickly. He means that the system's judgment holds steady, without sudden lapses.
Almeida's shorthand is that "you get Einstein like maybe 95% of the time and Mr. Bean the other 5%." Mr. Bean is the clumsy character from British comedy. The bad 5% wipes out the average; in expected value, he says, it just does not work out.
Almeida says he has watched company after company try to automate accounting over several hype cycles, and keep failing. He says the math does not work out, so companies learn they cannot actually automate the work.
The issue is what happens when software runs in the background. Almeida says he wants to write it once, put in effort up front, and have it "run in the background and solve the problem forever."
That kind of work needs nines of intelligent reliability, meaning correct behavior almost every time. For ordinary work, Almeida argues, trading a little genius for steadier judgment is the real superpower.
Questions people ask
What does 'nines of reliability' mean?
It is engineering shorthand for very high dependability, such as 99.9 percent correct operation, then 99.99 percent, and so on. Each extra nine means roughly ten times fewer failures.
Why not just have a human check the AI's accounting?
A person can catch mistakes, but then the job is not really automated. The company still pays for review, loses speed, and risks the one error that slips through. Almeida's point is that true automation means the software can run in the background without supervision.
Won't a smarter model automatically become more reliable?
Not necessarily. A system can become more impressive on hard problems while staying unpredictable on easy ones. Reliability comes from reducing surprising failures, which may require different tasks, data, and testing than raw problem-solving skill.
