Write the Story Around Your Numbers With AI

For years the worst hour of my month was the one where I turned a grid of numbers into English. Variance commentary. Everyone above me wanted it, nobody below me wanted to write it, and it never got easier, only faster. So the first time I watched a model do that hour in ten seconds, I should have been glad. Instead I sat there holding a draft I couldn’t send and no tidy way to say why.

Every line was true. That was the strange part. Software spend up 12%, headcount under plan, marketing heavy in the back half of the quarter. Each sentence checked out against the ledger, and each one was hollow. The machine had written the exact words I used to write, and standing next to them I finally saw what those words had always been hiding, which was nothing. I had spent years being paid, in part, for the typing.

That took a while to sit with. What I think was going on is this, and the day it clicked, the way I use the tool changed for good.

The draft has two layers, and the model only writes one

When I pulled a good piece of commentary apart, it came apart in two. One layer just restates the numbers in sentences instead of cells. The other says why it happened, what it does to the months ahead, and what someone should do about it. I had always written both in the same breath and never noticed the seam between them. The model found the seam for me, by doing one side flawlessly and leaving the other blank while writing as if it hadn’t.

It is genuinely good at the first layer. It can see that software spend rose 12% because the 12% is sitting right there in the file I handed it. It cannot see that the rise is three annual renewals that happened to stack into one month and won’t come back, because that isn’t in the spreadsheet. It’s in me. It’s the residue of years of sitting in the business. And when I ask the model for the cause anyway, it doesn’t tell me it doesn’t know. It invents one in the same calm voice it used for the true lines, and that is the whole trap.

I can follow how that goes wrong without having to watch it happen. The invented cause reads exactly like the real ones on either side of it, so it survives the skim, lands in the deck, and gets believed. Someone resets next quarter’s forecast to chase a trend that was really just renewal timing. My fluency, borrowed from a machine, becomes a wrong number in a plan I will have to defend later. Given that, I would rather leave the box blank.

So I stopped asking the model the one thing it can’t know. Now I give it the numbers and the audience and let it write the description, and I keep the why and the so-what for myself. It sounds almost too obvious once it’s written down. But watch how most people actually use these tools and they do it backwards, asking the machine for the insight and then editing its prose, which is handing over the only valuable half and lovingly polishing the one that was already free.

The description just restates the numbers; the judgment is the actual report.

Give it the boring layer, keep the valuable one

The prompt I use for the first pass is deliberately stupid. I want description and nothing else, so I say exactly that and shut the door on guessing:

Below is this month's actuals vs budget by line item. Write a plain
factual summary for a CFO who skims: one short sentence per material
variance, stating the line, the direction, and the size. Do not
speculate about causes. Do not recommend anything. Just describe
what moved.

[paste the variance table underneath this line]

What comes back is a clean scaffold, arithmetic stated, no story bolted on. Then I do the part that was always the actual job.

Back to that 12% software line. The model leaves it at “software spend rose 12% versus budget.” Mine says the rise is three renewals that stacked into one month, that the run-rate underneath is flat, that we shouldn’t touch the forecast over it, and that we ought to smooth renewal timing before next year’s budget so the same lump doesn’t spook us twice. None of that was in the file. All of it is the reason I still hold the job the model is doing half of.

One honest edge, and I would be lying to skip it. All of this assumes the numbers underneath are right, and the model is least trustworthy on exactly the derived figures commentary leans on. Describing them fluently doesn’t get me out of checking them. That is its own skill, the one about catching the hallucinated number before it hits the deck, and it comes first; then I narrate. The bigger pattern, letting AI take the mechanical layer of a task while I keep the judgment layer, is worth tracing across everything I do once I can see it, which is what auditing your job into tasks is for. It’s the same move as the assign-the-drafting, keep-the-decision half of the four ways to work with AI.

I don’t do any of this because I’m scared of the machine. I do it because it finally took the worst hour of my month off my hands and gave me back the one part I actually liked. Describing the numbers was never why I got into this. Working out what they mean still is. The tool writes the sentences now, and I get to go back to thinking, which is the trade I would have taken years ago if anyone had offered it.