Skip to content
Go back

Why AI Fails at Invoices but Excels in Surgery


Recently, I came across a real case that illustrates—almost ironically—the limits of artificial intelligence in everyday business.

A company decided to automate part of its accounting workflow using an AI model based on ChatGPT to process invoices in several of its branches. At first, everything looked promising: the system handled documents quickly, employees celebrated how manual work was reduced, and the experiment seemed like a success.

But soon problems started to emerge: invoices with incorrect addresses, misinterpreted fields, duplicated entries. What initially seemed like a breakthrough ended up causing more than $10,000 in losses.

Meanwhile, in another part of the world, thousands of complex surgeries continue to be performed with the Da Vinci robotic system, renowned for its millimetric precision and high success rates.

So here’s the paradox:
Why does AI stumble over something as “simple” as an invoice, yet can assist in something as delicate as surgery?


The paradox of “simple” vs. “complex”

The answer lies in the difference between closed, consistent environments and open, chaotic ones.

In the operating room: a closed, controlled world

In short: the operating room is a closed, predictable environment, and that’s why robotic assistance can be so reliable.

In the accounting office: an open, messy world

That’s why errors pile up: the AI isn’t “understanding,” it’s guessing statistical patterns.


AI is statistics, not thought

Mathematician Terence Tao puts it bluntly: AI doesn’t think—it’s statistics. It doesn’t “understand” what it does. Instead, it adjusts millions (or billions) of parameters to predict the most probable output.

Ironically, what looks harder (operating on a human body) is actually more constrained than what looks easier (processing invoices).


The lesson for businesses

The takeaway is clear:

If we want AI to handle invoices with the same reliability as robotic surgery, we first need to redesign the system around it: business rules, human validation, standardized data pipelines.


Final reflection

Perhaps the real question isn’t “What can AI do for me?” but “What conditions must I create for AI to actually work well?”

The difference between a $10,000 accounting loss and a successful life-saving surgery is not in the “intelligence” of the machine, but in the degree of order, control, and consistency of the environment where it operates.


Share this post on:

Previous Post
AlphaFold doesn’t predict words, it predicts shapes
Next Post
Can an AI Be Conscious? A Centrist Manifesto for Careful Thinking