My AI disagreed with my EBITDA. It wasn't wrong.

Hey {{first_name|there}},

My AI dashboard showed an EBITDA I didn't recognize.

Not by a lot. But it wasn't mine. Same with EBITDA margin and a few other lines.

Here's how I work. My team builds the first version manually. A management report, for example. Then I give AI the same job and compare.

If the numbers match, I trust the process. If they don't, I find out why.

This time they didn't match.

You've heard "garbage in, garbage out." This isn't that. The AI did exactly what I told it.

This is email 2 of 6. In email 1, I promised to explain why clean data mattered more than the AI itself. What actually happened is more specific than that.

What I found

I asked the AI where each number came from. I checked its answers against our files and they held up.

Three things:

1. A definition:

Before building anything, the AI asked which adjustments I'd make to EBITDA. I answered naively. I didn't think about what it would do with my answer. It used them right away and went straight to adjusted EBITDA.

In our reporting, we calculate EBITDA first. Then we apply the adjustments and show the bridge to adjusted EBITDA, with the reason each one is non-recurring or non-operating.

The AI skipped the first step. Two different numbers that sound like one.

2. Missing context:

The AI didn't know how my business is structured, so it couldn't tell which costs are variable. Without that, it couldn't know what contribution margin should include.

3. Two budget numbers:

The AI flagged two budget figures for the same line. I traced both back to their files. Two versions of the budget, with a small difference between them.

The approved one lives in our FP&A system, and it was right. But a number from the previous version had been used in places, so different numbers were showing up in different files.

The AI wasn't wrong in any of the three. I hadn't told it what my words mean, and our files didn't make clear which budget was final.

What I changed, and what it bought me

We wrote down a definition for every KPI: the formula, what's in, what's out, and why.

We described how the business is structured, including which costs are variable.

We gave the AI examples of where each number comes from.

And I made sure every report used the final approved budget.

My monthly reporting used to take 5 working days. Now the clock starts when my team's package lands with me.

I gather it and put it in the system.

From there, it takes about 2 to 3.5 hours, depending on data volume and how many questions the AI asks me. Three cycles in a row and the results are impressive.

That mismatch is part of how I got there. The AI could only be fast once it knew what my numbers meant.

Two habits worth taking from this

When a number looks off, ask the AI where it got it. Then check the answer against your files.

When it asks you questions, answer like you're giving instructions. It will use every word.

Thursday, email 3 of 6: the executive deck I've never been good at building by hand, and what AI built instead.

One question:

Which KPI do you wish everyone in your company defined the same way?

Hit reply. I read every answer.