Auditing and reviewing financial model data, assumptions, and consistency

I was a panelist on a panel about AI with Naturally NorCal yesterday. Here is what I learned while teaching.

I have had a thought bouncing around my head for a long while. Yesterday, it finally crystallized. A 2x2 came together in my head to help map out how we should actually be using these tools.

On one vector is my level of expertise. In some areas, like CPG financial operations, I am an expert. In others, I am less than a novice.

On the other vector is the criticality of the output. I use AI for strategic, foundational projects at Cultivar, but I also use it to make cartoons with my kids.

Here is how those quadrants break down:

1. Expert + Critical Output: When I know the subject matter deeply and the work matters, AI helps me 10x my productivity. It handles the heavy lifting so I can focus on the strategic edge.

2. Novice + Critical Output: This is where AI is dangerous. If you do not have the expertise to spot a hallucination or a systemic error in a complex P&L, AI alone is not sufficient. You still need to call in outside experts.

3. Novice + Non-Critical Output: When the stakes are low, AI is a wonderful way to play. For me, this is helping my kids unleash their imaginations by creating books and cartoons.

4. Expert + Non-Critical Output: When I am an expert but the output is not my primary focus, I find that using other experts' tools is the most productive use of time. For example, the team at Glimpse is helping us scale our trade spend management service line with their AI-powered deduction dispute tools.

We are all trying to find that balance between precision and speed. How are you using AI in your daily ops?
----
Hi, I’m Pedro, CEO of Cultivar. We help CPG Founders feel confident about their finances by managing accounting, inventory, trade spend, and bev alc compliance so you can focus on growing the brand.

Investment Banking: Ranking the Best AI Tools for Financial Modeling (2026) 🤯

We extensively tested OpenAI's ChatGPT, Anthropic's Claude, Microsoft Copilot Agent Mode, and Shortcut on a real three-statement Excel model using investment banking standards.

Here’s what actually works, and what doesn’t.

Quick Answer: The Rankings 👇

Best Overall: Shortcut

Close Second: Claude (Opus 4.6)

Third Place: Microsoft Copilot (GPT-5)

Distant Fourth: ChatGPT (GPT-5.2)

The Testing:

Criteria 1: How It Feels Working With The Tool

Understanding the Assignment:

Winners: Claude and Shortcut

Claude and Shortcut asked thoughtful clarifying questions after receiving the prompt, about:

• Forecast preferences
• Revenue segmentation
• Share repurchases
• Layout decisions
• Schedule structure

That behavior closely resembles what you’d want from a good junior analyst.

Copilot and ChatGPT asked none.

Speed:

Winner: Shortcut

Shortcut and Claude completed the setup in ~15 minutes vs ~25 minutes for Copilot.

ChatGPT took close to an hour.

A decent analyst would have taken 1-2 hours to complete the assignment.

So all were faster than a human analyst at initial setup.

Criteria 2: Data Extraction, Formatting and Best Practices

Formatting:

Winner: Shortcut

Shortcut and Claude produced the most “investment-bank-like” outputs.

Shortcut was more consistent with input coloring and structure.

Claude missed several formatting conventions.

Copilot ignored IB formatting entirely.

Accuracy:

Winner: Copilot

Copilot won, but this was disappointing across the board.

Shortcut and Claude hallucinated significant portions of historical data.

In both cases, the errors were subtle enough to be dangerous, with slightly incorrect line items all adding up to correct subtotals.

Shortcut’s second attempt returned almost no mistakes.

Claude continued to generate bad data.

Fixing this would require careful cell-by-cell auditing that takes longer than just inputting the numbers yourself.

As a rule, analysts should not rely on these agents to find data and should instead upload PDFs and spreadsheets for the agents to work with.

Copilot and ChatGPT were more accurate.

ChatGPT’s presentation was the least polished, but its historical balance sheet was easiest to audit.

Had Shortcut and Claude used correct data, they would have won as they were attempting a more analytically rigorous presentation.

Shortcut was also going into the footnotes to break out certain items when appropriate.

Sourcing and Commenting:

Winner: Claude

Claude provided the best explanations of where data came from and why certain modeling decisions were made.

Copilot added no comments.

ChatGPT added too many.

Shortcut did some, but less consistently.

Claude was also the only tool to backsolve EBITDA correctly.

The Bottom Line:

Shortcut and Claude significantly outperform Copilot and ChatGPT.

But right now, even the best tool still underperforms a Junior Analyst.