Last month, I was training a big American Saas company.

In this session, I showed 35 of their finance pros two AI analyses of the same cost center data.

The first analysis completed really, fast.

The second analysis was much slower.

Both of them said "6% year-over-year OPEX growth."

So I asked: "Who knows this difference?"

Almost nobody.

Here is the secret - AI has two modes.

And the difference is not just about your accuracy. It's about your quality, and it’s about your audit trail.

Because, If you're using ChatGPT in Instant (Fast) mode for financial analysis, you're only going to get basic analysis.

As an ex-auditor, I can tell you that detail means everything.

And, as much as using AI super fast might give you some basic calculations.

You need to ask yourself whether this is something you can defend during the audit.

If it’s not, this is causing for you a big problem.

But now you know this problem exists, I'm going to show you exactly how to fix it.

Generative Analysis (Bad vs Good)

With AI, there are now standard AI models, and reasoning (thinking) AI models.

But, because reasoning models take more time to run, a lot of tools default to the standard model.

So, when you upload your P&L and ask for analysis in instant mode, AI sometimes gives you simple math and basic results.

It's not that it can't calculate, it's that it doesn't think hard enough about the analysis.

This means, it is more likely to be inaccurate, and more likely to create gaps in your audit trail.

But, in thinking mode, instead of just calculating, it analyzes.

Instead of giving you the answer straight away, it does things like:

• Checking data consistency first (do the totals reconcile?)
• Proposing analysis types before running them
• Performing calculations in Python (so you can audit the code)
• Creating visualizations for each analysis
• Writing commentary explaining what it found

This also provides a much more detailed audit trail as well.

You can see every formula. So, if it says `average by project` when it should be `sum by project`, you catch it and fix it (before the board meeting).

Most finance teams are using instant mode for everything. Because they don't know thinking mode exists.

Now you do.

How to Force AI to Calculate (Not Guess)

Here's how you switch to thinking mode, improve your analysis, and verify your numbers.

Step 1: Always Start with Thinking Mode

When you upload financial data to ChatGPT Business or Enterprise, look for the mode selector at the top left.

Depending on your plan you will see different options:

• Auto (where ChatGPT decides - default)
• Instant (the fastest least accurate response)
• Thinking (this is the calculator mode that you want)
• Pro (Research grade - only available on some plans)

Click Thinking before you attach your file.

(In some plans you will also have access to ‘Standard’ thinking and ‘Extended’ thinking depending on how complex the task is)

• For Gemini, it's called "2.5 Pro" (not "Flash").
• For Copilot, use the "Analyst" agent as it has thinking mode built in.
• For Claude, click the ‘Extended Thinking’ button before you chat

Pro Tip - If I am working on something that does not need thinking mode when I’m using ChatGPT, I press the ‘Skip’ button, and it generates the output straight away.

Step 2: Use the 5-Step Analysis Prompt

Don't just say "analyze this file." That's what gets you garbage.

Instead, give AI the same instructions you'd give a junior analyst. Here's the prompt I use:

This comes from my auditor background. You always check data consistency first. Because if your totals don't reconcile, nothing else matters.

Step 3: Look for the Python Code

Before you accept any calculation, check if AI actually computed it.

Click "View Analysis" or "Show Code" (depending on which tool you're using).

You should see something like this:

Step 4: Audit

Now, you are probably not a Python developer.

But you still need to audit it for accuracy, so:

1. Copy the code
2. Start a new chat
3. Type this prompt:

You will then get a breakdown of the calculation in a format that you can audit properly.

​

​

The Bottom Line

You can't trust AI to calculate any more than you'd trust autocorrect to file your taxes. But you can trust AI with a calculator, but only if you force it to use one.

Thinking mode is about making AI show its work so you can act as an auditor, not the co-conspirator in numbers that just "look about right."

Most finance teams are using instant mode for everything because they didn't know there was an alternative (or they just let ChatGPT choose for them).

Now you do. So, next time you upload financial data, switch to thinking mode first. Then spot-check one calculation manually. If the Python code shows real formulas and the math is correct, you've got a junior analyst who works in minutes instead of hours.

If there's no code? You've got a problem you just caught before the board meeting.

But you now know how to fix this.

And your AI accuracy just got 10x better.

​

Best,

Your AI Finance Expert,

Nicolas

Behind The Scenes

Last month I presented at Forge Connect (Antwerp) alongside my friend Soufyan Hamid.

My presentation was on AI-powered data cleaning and analysis, and Soufyan presented impactful storytelling in finance.

P.S. - What did you think of this edition? Hit reply and let me know (I read all replies).

P.P.S. - In this video I show you the difference between prompting using different modes and how reasoning or 'thinking' mode is best for analysis. Watch here!​