Let that sit for a second.

This is the problem nobody in FP&A wants to talk about: AI can find patterns faster than any analyst alive, but it can’t tell you if those patterns mean anything.

Why This Matters Right Now

Everyone’s rushing to plug AI into their forecasts, their variance analysis, their planning cycles. And I get it. I’ve done it. I rebuilt parts of my own close workflow using Claude, and the time savings are real.

But here’s what I’ve noticed in the conversations I’m having (and the LinkedIn posts I’m reading): there’s a growing assumption that if the model says it, it must be true. That a high R-squared or a strong correlation is the same thing as an explanation.

It’s not. And in FP&A, where your variance commentary lands on the CFO’s desk and shapes real decisions, confident nonsense is worse than no answer at all.

Correlation Is the Easiest Thing in the World to Find

Tyler Vigen built an entire website (Spurious Correlations) and a book around this idea. He took thousands of variables and let the math find connections. Margarine and divorce. Swimming pool drownings and Nicolas Cage films. Per capita cheese consumption and the number of people who died tangled in their bedsheets.

Every one of those correlations is real. The math checks out.

And every one of them is meaningless.

The reason is simple: machine learning algorithms (and even basic regression models) are pattern-matching engines. Give them enough variables and enough time, and they’ll find relationships. That’s what they’re designed to do. But they have zero ability to distinguish between “these two things move together because one drives the other” and “these two things happened to trend in the same direction for a decade.”

I’ll be honest, I caught myself falling into this trap early on. I was experimenting with an AI-assisted variance analysis workflow and the model surfaced a correlation between a GL line item and a completely unrelated operational metric. It looked compelling in the output. Clean chart. Strong relationship. I almost included it in my commentary before I stopped and asked myself: can I actually explain the mechanism here?

I couldn’t. Because there wasn’t one.

The Real Danger: You Can’t Explain a Variance by Pointing to the Algorithm

This is where it gets practical.

In FP&A, when you present a variance bridge or a forecast miss, the question from leadership is always: why? Not “what does the model say?” but why did this happen and what do we do about it?

If your answer is “the algorithm identified this as a key driver,” you’ve given them nothing they can act on. You’ve just dressed up a correlation in a suit and hoped nobody asks follow-up questions. (They will.)

The job isn’t to find patterns. The job is to explain the business.

And that requires understanding the data inputs you’re feeding the machine. Every feature in your model carries assumptions. Every variable you include is a hypothesis about what drives the outcome. If you don’t scrutinize those inputs with the same rigor you’d apply to a manual analysis, you’re just automating the process of being wrong faster.

Quick note: all the numbers in this article are illustrative. I don’t publish proprietary data from my employer. The patterns come from real work. The specific dollar amounts are made up.

Here’s a practical example. Say you’re building a model to predict monthly SaaS churn. You include 15 features: usage metrics, contract terms, support tickets, NPS scores, industry segment, deal size. The model runs and tells you that “number of support tickets in the last 90 days” is the strongest predictor of churn.

That might be true. Or it might be that customers who are already planning to leave stop engaging with support entirely, and the customers who file lots of tickets are actually the ones fighting to make the product work. Same data point. Opposite causal story.

You don’t figure that out by reading the model output. You figure it out by talking to the CS team, pulling the cohort data, and looking at what happened after the tickets were filed. That’s the analyst’s job. The AI can’t do that part.

What I Actually Do Differently Now

After getting burned a few times (nothing catastrophic, but enough to make me careful), I changed my workflow. Here’s the short version:

Before I let AI touch a variance or a forecast driver, I answer three questions manually:

1. What is the business mechanism? Can I draw a straight line from this input to the outcome that a non-finance person would understand? If I can’t explain it in one sentence, it doesn’t go in the model.

2. Would this relationship hold in a different time period? If the correlation only exists in the specific window I’m looking at, that’s a red flag. (This is exactly what the margarine/divorce example illustrates. Zoom out, and the relationship disappears.)

3. What would change if I removed this variable? If the model’s output barely moves, the variable was noise. If the output changes dramatically and I can’t explain why, I need to dig deeper before trusting it.

This adds maybe 15 minutes to the front end of any AI-assisted analysis. And it’s saved me from presenting correlations that would’ve eroded my credibility the first time someone asked “but why?”

The Mental Model: Think “Features, Not Answers”

Here’s the frame I keep coming back to: AI gives you features. You supply the meaning.

A feature is just a variable the model uses. “Margarine consumption” is a feature. “Divorce rate” is a feature. The algorithm doesn’t know or care that these are absurd things to connect. It just knows they move together.

Your job as the analyst is to be the quality gate between the model’s output and the story that reaches leadership. That means interrogating every feature, every input, every assumption before you trust the result.

Not after. Before.

This isn’t anti-AI. I use AI in my workflows every week. But I use it the way I’d use a very fast, very confident junior analyst who has no business context: I check the work.

What to Do This Week

• Audit one AI output you’ve recently trusted. Pull up the inputs. Can you explain the business mechanism behind every variable? If not, flag it.

• Add a “causation check” to your AI workflow. Before any model result goes into a presentation or a forecast, force yourself to answer: “Can I explain why this relationship exists?”

• Read Tyler Vigen’s Spurious Correlations site (tylervigen.com/spurious-correlations). Spend 10 minutes. It’s funny, and it’ll permanently change how you look at model outputs.

• Talk to the business. The best validation for any data-driven insight isn’t a statistical test. It’s a 10-minute conversation with someone who lives in the process every day.

The analysts who get ahead with AI won’t be the ones who adopt the fastest.

They’ll be the ones who know when to trust the output and when to ask harder questions.

That’s the difference between using AI and being used by it.

And that’s all for today.