|
So, your analyst is sending you the monthly variance commentary.
It's well-written and insightful. It even identifies the key drivers perfectly.
You're impressed.
Until you check the numbers.
Revenue variance? AI said +12%. The real number was +8.4%.
That "concerning pattern" in SG&A? It was completely made up. The writing was perfect, but the analysis was wrong.
So now you have this choice:
- A) You can review every AI output line-by-line.
- B) You can ban AI entirely and lose all the efficiency gains.
Neither option works.
So here’s a better way.
AI Audit
|
|
I spent 7 years as an auditor for PwC. And, I tell everyone in my training sessions the same thing…
You cannot hide behind AI.
You own the work you produce.
When it's really good, you should own it and be super proud. But when it's bad, you should also own it and make sure you correct it so that it doesn't happen anymore.
And this is something really important to understand. Generative AI doesn't calculate. It only generates information. If you don't see Python code or if you don't have an Excel file with formulas, this is a red flag. You need to make sure that it has been calculated somewhere.
The step one is always to verify the consistency of data. Because as an auditor, if you didn't do that and you started your test and the total didn't match the financial statements or the trial balance, basically you had the wrong file and you should not even start working.
This same logic applies to AI. You don't verify everything. You verify the right things in the right order.
Fast AI Audit Process
Step 1: Check Data Consistency First
|
|
Before AI analyzes anything, verify your source data totals reconcile. This is basic, but I cannot tell you how many times I've seen people skip this.
Upload your file and ask AI:
"Check if the detail lines add up to the totals. Flag any differences."
AI is actually really good at this. It will find mismatches between tabs, flag rounding differences, and even ask you whether you want to work from the details or the totals. But the point is: you catch the problem before you start, not after you've built an entire analysis on bad data.
If data doesn't reconcile, stop. Fix it first. Don't do analysis on bad data.
Step 2: See the Calculation
|
|
Whenever you use AI to do analysis or calculations, you need to use reasoning models that think through the process. You can read my previous newsletter where I show you the difference here.
When you use a ‘thinking’ model, the AI will use Python to do it’s calculations.
|
|
When you see Python code, and it says something like "group by country" and shows the totals for France and the US being grouped with that formula, you know it was calculated and not just generated by AI.
I ask AI to show me the calculation. Because like this, I can audit it really quickly. Having 1.6 versus 1.563 in front of me, I can see okay, yes, the difference is 73 out of 1.5, that's about 5 percent. I can make my spot check because I asked to show me the calculation.
Step 3: Make AI Audit Itself
Think of AI like a colleague. You would not just accept a colleague's work without asking them to document the checks they did before you start reviewing. Same thing here.
One of my members, Manny, asked a really smart question about whether you can use another LLM (Large Language Model – ChatGPT, Gemini, Claude, Copilot) to verify.
But, you don't even need another LLM. Just copy/paste the results in another conversation. You can even do this in the same conversation. Just ask AI to review its work and see where it could have gone wrong.
Ask it to show you how it audited the work.
|
|
AI's self-review often catches its own hallucinations.
Step 4: Trust Your Radar
You have a lot of experience, and you can use this to use AI much better.
You probably will pick out problems much faster than somebody who’s only just started to enter their finance career. You have built a radar for this sort of thing.
You know what type of analysis are the right ones. You know the language. You know where the data should link. So when something doesn’t feel right, trust that instinct and make sure to review.
So here is what I really want you to take away from this.
Do not trust AI blindly. The output is your responsibility. Your name is on it. Your credibility is on the line. That does not change just because a machine helped you write it.
But, with the right approach you can absolutely use AI for financial analysis. If you know how to audit it smartly, it becomes one of the most powerful tools in your career.
Don't be left behind because of your fear. Use it to be smarter than the rest ;).
The One Thing To Remember
Next time someone on your team (or you) uses AI for analysis, add one step before you accept any numbers. Ask AI to show the calculation and review its own work.
It takes 30 seconds. And you'll immediately see whether you're looking at real analysis or something that has been convincingly made up.
AI is the tool. But you've still got to own the output.
This is the difference between an AI Finance Beginner and an AI Finance Pro.
Best,
Your AI Finance Expert,
Nicolas
P.S. – Did you try this? Hit reply and let me know what you found. I read all replies.
P.P.S. – If you want to see more of what AI can do, I have covered three practical AI use cases in one of my videos. You can check out here 3 Practical Ways to Use AI as a Finance Pro in 2026