Why AI Fails For Operators (And How To Fix It)

TL;DR

Everyone treats artificial intelligence like magic. It is just software. The real barrier to using it is not technical, but operational: operators look at a blank screen and have no idea what to ask. This piece breaks down exactly what AI can do, what it cannot do, and how you can use it to test if your thinking is actually original.

The Blank Screen Problem

Artificial intelligence fails to stick because most people look at a blank chat box and have no idea what to ask.

When spreadsheets first appeared in the late 1970s, you had to buy a computer worth thousands of dollars just to run them. But if you showed VisiCalc to an accountant, they instantly understood it. They saw the rows and columns. They changed one interest rate, and the entire board updated. It took a week of work and compressed it into thirty minutes.

We do not have that clear mental map for language models yet. You open a text box, and nothing guides you. I am still figuring out how to map my daily operations to what the machine is actually good at doing. If the machine does not come wrapped in a specific product—like a button that says “draft a reply to this angry client”—you have to invent the use case yourself. Most people only think of a new use case once a week.

Track exactly which repetitive tasks eat your week and build one specific prompt to handle just one of them.

The Limitation of Being Roughly Right

Large language models are statistical systems, which means they are useless for quantitative analysis where exactness matters.

They generate answers that are roughly right. Roughly is a spectrum. If you ask a machine to write a professional biography for an event, it might get your graduation year wrong. You can read the draft, spot the error, and fix it in thirty seconds. That saves you an hour of staring at a blank page. The tool proved useful.

But if you ask the machine to run your cash flow model, roughly right is dangerous. You do not want a system that hallucinated a number three times on a single page managing your working capital. We do not currently have a model that knows when it is mathematically wrong.

Only delegate tasks to AI where the cost of verifying the output is lower than the cost of doing the work from scratch.

Raising the Baseline of Insight

Artificial intelligence serves as a perfect filter for unoriginal thinking.

Writing is the mechanism we use to figure out what we think. Before, you would write something and ask yourself if you were adding value. Now, you have a much harsher test available. You can write your draft, look at it, and ask: is this exactly what an AI would have generated?

If the answer is yes, you have not actually added any value. The machine sets a new baseline for what qualifies as an insight. If a language model can produce your exact thought from a simple prompt, anyone could have said it. It forces you to throw out the obvious and ask the next, harder question.

Before publishing any operational update or field note, paste it into an AI and delete anything the machine could have written on its own.

Questions to Consider

1. What is one specific, recurring task where a “roughly right” draft saves you time?
2. Where are you asking AI to do exact math, and how can you move that back to a spreadsheet?
3. Which of your current business processes rely on a blank chat interface rather than a fixed system?
4. If you run your latest operational thesis through a model, does it sound identical to the machine’s output?

Quotes

“You can write your draft, look at it, and ask: is this exactly what an AI would have generated? If the answer is yes, you have not actually added any value.”

“Only delegate tasks to AI where the cost of verifying the output is lower than the cost of doing the work from scratch.”

“Artificial intelligence fails to stick because most people look at a blank chat box and have no idea what to ask.”