Regular readers know I’ve been a vocal fan of Claude for Excel for a while now. It’s been my go-to for spreadsheet-based financial analysis and I’ve covered it extensively in this newsletter. So when OpenAI dropped their own version, I wanted to see how it stacks up.
This isn’t a casual release either. OpenAI has invested heavily into this product. The expectations coming in were high.
(I wrote about my ChatGPT Excel predictions in October: Link)
Coming after excel work has been in the works for at least 6 months at OpenAI. My prediction about this from October: Link
For most people, an AI Excel plugin sounds like a quality of life upgrade.
For investors, this is something bigger.
Building a full 3-statement model from scratch is one of the most time-consuming tasks in fundamental analysis. Sourcing documents, cleaning assumptions, wiring cash flows… a proper build on a large cap can eat an entire day. And for fixed income and credit investors specifically, getting the cash flow mechanics right is non-negotiable.
So for this week I wanted to test this new Excel plugin using a fixed income workflow.
Specifically: building a full cash flow focused 3-statement model for META.
Here’s the plan:
Use Perplexity Computer to scrape all relevant company documents from the last 3 years (IR presentations, SEC filings, financials)
Run a prompt to compile a structured assumptions pack
Upload that assumptions pack into ChatGPT for Excel via RAG to build the model
The Prompt
We are using 3 prompts to do this task. First one prepares the data room (via Perplexity Computer). Second one extracts the model assumptions pack (via the same Perplexity Computer instance). And the third prompt takes this assumptions pack to build the model.
Prompt 1: Use Perplexity Computer to Prep Data Room
Verdict: I found several errors on ChatGPT for Excel's modeling mechanics.
The first pass looked impressive. Visually clean, statements appeared to tie, and the model populated faster than any manual build I’ve done. But here’s the thing about financial models… looks can be deceiving.
I ran a full audit pass on the output and found issues that would materially change any capital decision made off this model.
The most critical problem was in the first forecast-year working capital build. That section was mis-linked, with formulas referencing the wrong prior-year lines. The clearest example was in the accrued-expense bridge, where the model pulled the prior-year accrued change line instead of the prior-year accrued balance. This error materially overstated FY2026 cash generation.
On my review, the effect was roughly:
FY2026 CFO overstated by about $24.6B
FY2026 FCF shown at about $35.8B instead of roughly $11.3B
FY2026 net cash / (debt) shown at about +$15.0B instead of roughly -$9.5B
A credit investor making a decision off that first pass would have a fundamentally wrong picture of META’s liquidity position.
There were other issues as well:
Historical net cash was wrong because the model subtracted long-term income taxes instead of long-term debt
The balance sheet check was not a robust control because forecast equity was a balancing plug
To be fair to the plugin: all the core historical hardcodes I verified against META’s FY2025 10-K matched cleanly. Revenue, operating income, net income, CFO, D&A, buybacks… all correct. The bones were right. The formula mechanics had holes.
The good news is you can self-correct. I went back and forth prompting ChatGPT to audit its own model and fix the mechanics, uploading the workbook directly into the ChatGPT front end for the audit pass. It found and documented the errors itself with enough of the right prompting.
The bottom line: ChatGPT for Excel is a genuinely strong first step. The workflow works. But treat v1 output as a draft, not a deliverable. You need a second pass audit prompt before trusting any number for a real capital decision.
The edge here is still massive… what used to take a full day now takes minutes plus an audit loop.
(I'm confident OpenAI will improve this plugin as they get more training data...let's check back in in a few weeks...)
For now: just don’t skip the audit step and blindly trust the v1 of the model build. I'm sticking to Claude's plugin until OpenAI improves the model mechanics error rate.
Personal
If you want to go deeper on the $META thesis, I wrote an article on my for why I'm a $META bull.
The TLDR of my investment thesis:
App creation is exploding via agentic coding. All these apps need somewhere to spend to acquire users which causes ad bids for META and GOOG
The new Chinese video models are incredible and only social media platforms protected by Section 230 can host these entertaining deepfake videos ... I bet platforms like Instagram take eyeball share away from Netflix
Meta's AI ad machine is already compounding fast (driving more ad dollars to the platform)
IMO the institutional demand for inference does actually materialize which will be a new massive business unit
Llama gives them model optionality nobody is pricing in. Similar to Google 2023 when everyone wrote off Gemini.
WhatsApp monetization is a sleeper revenue line.
Taking a step back, I will say I've noticed a massive difference in my investing process (and returns) since I've transitioned to an AI-native framework these last two years. If you haven't made the shift yet, I highly recommend investing the time!