These CFOs spotted 15 unusual P&L outliers in 4-steps (run this analysis now)

What, Why, So, Now?

Today I’m going to show you how the CFOs in my community use AI to find outliers / anomalies.

This is what happens when you as AI to help you with this:

1. A machine (normally Python that AI runs in the background) looks at your history and works out what is seasonal and what is unusual.
2. A language model reads that output and explains it. The maths is done by the machine. The story is written by the AI model.

What’s super important to know here is that the model's story can be wrong even when the maths is right. So this is where your human skills come in.

Once the AI has surfaced the outliers, you A) decide whether you agree with the story, and B) take it a step further.

1. What happened? The plain fact. This account moved by this much, in this month.
2. Why did it happen? The likely driver. Timing, a one-off, or a real operational change.
3. So what? Does this actually matter to the business, or is it just noise?
4. Now what? The decision, or the question you take to the stakeholder.

The AI can draft 1 & 2. But 3 & 4 require your narrative and insight.

Two weeks ago I set this as an ‘AI Workout’ to our AI Finance Club members. We give them a real task to use with their data, and they report back (with the data anonymized)

The results were awesome. Marianne, a fractional CFO, turned her output into a one-page exec "Unusual Movements Pack" with the flagged lines, the variance size, and the exact owner question per line.

Chris, another fractional CFO, ran it on live company data and said it surfaced insights that had not yet been discussed with company leadership.

If you can find the data nobody else has found yet, you can build yourself a very strong position.

Here’s how to do it.

​

​

The 4-step Unusual Movements Workout

By the end you will have a ranked list of your top 15 anomalies, five questions for your stakeholders, and a one-paragraph story for leadership. Here is the exact workout.

1. Check and use clean data

Use your monthly P&L by account or cost centre, last 12 to 24 months. But strip it to one amount column first. I have seen ERP exports with ten columns for the same number, and AI often uses the wrong one.

Make sure to keep individual payroll lines out. Just aggregate them at team level, and check the totals add-up before you start.

2. Give it the workout prompt

With your clean data, ask:

3. Use What, Why, So and Now

Take each anomaly through what happened, why, so what, now what. Where the "why" is unclear, you can do a quick fishbone analysis or five-whys pass to get to the real driver (you can use AI to help you for this as well.

This is the step that stops you taking a wrong number into a board meeting.

4. [Advanced] Make it prove the outliers in Python.

Ask it to run a real statistical test using Python. Some examples:

• Z-score – measures how many standard deviations a number sits from the average, so anything beyond about 3 is a statistical outlier rather than normal variation.
• IQR fence – ranks your data, takes the middle 50%, and flags anything sitting well outside that band; good for messy real-world numbers because a few extremes don't distort it.
• Isolation Forest – a machine-learning method that spots the points hardest to "box in" with the rest of the data, useful when an outlier only shows up across several variables at once (e.g. this cost is fine, this month is fine, but this cost in this month isn't).
• Box plot – the chart that shows all of the above at a glance: the middle of your data as a box, the normal range as whiskers, and outliers as individual dots.

(again, you can ask AI to suggest the best type of analysis based on the data)

The good news is, most AI tools now run Python right in the chat, so you don’t need to use a dedicated tool like Google Colab to run Python. So you can try some of the above and see what output AI gives you.

Just use more advanced tools if you want to keep the code the same, and be able to run it consistently every month.

For more information on how to use Python for advanced analysis you can read my previous newsletter ‘Machine learning in 6 steps: How to make better decisions with finance data your brain can’t see’ here.

Then at the end, you can do what Marianne did. Turn it into the one-page Unusual Movements Pack and present it.

The One Thing to Remember

Finance has always been a fight between speed and accuracy.

A few years ago, the faster a human worked, the more errors they’d produce, the more work for management in tidying up the mistakes.

But now, AI allows us to be faster AND more accurate. Providing we focus more on our human intuition when interpreting the numbers.

So use AI as a smart analyst, whilst you decide on the story.

You’ve now more opportunities than ever to ‘think outside of the box plot’ 😉

Best,

Your AI Finance Expert,

– Nicolas

​

P.S. – The best way to learn from me directly is to join my weekly masterclasses. 60-mins + Q&A where I’ll get all your AI in Finance questions answered. Join me here.​

P.P.S – Want to see the full outlier analysis run live? Here's how to do it in 5 minutes → How I Perform a Financial Analysis With AI in 5 minutes​

​