After getting acclimated with Anthropic's Claude for the last few weeks and 'vibecoding' a personal project on their free plan, I ponied up for the max plan and hooked into 'Claude in Excel.' If you’re thinking about firing your analyst, you may want to read the entire post!

The Scope: Build a Financial Model With Claude in Excel

I told Claude to build a commercial proforma model. I wanted a standard commercial proforma with a basic debt and equity capital stack, rent inputs for commercial tenants, monthly cash flows that roll into an annual cash flow, a returns summary, and an assumptions tab.

Note: I was much more specific than I'm describing here, but I won't bore you with the details. It's important your prompts are dialed in, but that's a discussion for another day.

The AI did its thing while I stepped away from the computer, and when I came back, my jaw hit the floor.

I thought, "This is the END of financial analysts as we know it!"

It built a FLAWLESS proforma model in minutes.

But then I started to REALLY look.

AI’s Mistakes and Oversights

There were some massive blunders, including:

The principal paydown of the loan was being added to the cash flow after debt, overstating the cash flow
The loan payoff in Year 10 was being added to the final equity distribution (not subtracted from it)

I found these mistakes immediately because the IRR and Equity multiple were so obnoxiously high, it was obvious something went awry.

But there were also minor mistakes that could have easily been overlooked

The 'Interest Only' toggle in the financing section didn't work
The cash-on-cash return formula logic was bad and inaccurate
“Other Income” was randomly a cell input on the “Monthly Cash Flow” tab (should have been on the “Assumptions” tab with the others)
I generally hated how the data was displayed (not a mistake, but annoying)

Regarding the last bullet, when I prompted for some layout/design changes, it killed the entire model.

It sure looks nice, though, beyond all the #REF! errors, doesn’t it?

And all I got from Claude was a "Sorry, I shouldn't have done that. Hit Control-Z to undo this. If that doesn't work, start over."

Claude apologizing for breaking my model.

It didn't work...Thankfully, I had just saved the file.

Tip: Anytime you are about to implement a big change with Claude, SAVE your work beforehand.

I'm still toying with the model, but it got to a point where I basically stopped using the AI altogether.

For example, I asked Claude to move some data higher up on the sheet (not alter cells or do any calculations, just move it up). It took forever to complete this task (and burned way too many credits), and it didn't even do it correctly.

To do this manually, I just had to grab the range, select Shift, and drag it up.

I basically got to the point during fine-tuning where I forgot Claude existed. AI was slowing me down at this juncture.

AI’s Impact on CRE Modeling

Claude in Excel is a tremendous assistant IF you know what you're doing.

It will absolutely save me countless hours, especially on the 'basics,' but I will need to babysit it, especially as the scope gets more niche and complex.

I envision using AI assistance to build the foundation of a proforma model. As I add greater complexity, I will lean on AI to help me do so more efficiently and to vet my own thought process in very small steps. I will never prompt multiple alterations/additions at once. I will prompt for one change at a time and verify its accuracy immediately after.

Fine-tuning will be done almost exclusively manually (at least for now).

Note: I also have some room for improvement in my prompting and being clearer and more direct when giving my instructions.

I think the most exciting part is that Claude in Excel is the worst it is ever going to be, and I was still amazed by some of the feats it could pull off in its infancy.

Using AI on Daily CRE Workload

I shared this analysis with my email subscribers, and I got a tremendous amount of responses asking to do more AI implementation on the everyday grind of CRE analysis. In other words, most groups already have sophisticated proforma models they’ve spent years on perfecting, and the true value of AI would be cleaning data, populating models, conducting market analysis, and streamlining the more tedious processes to populate a detailed proforma model.

Those of you using AI agents in Excel, what are you seeing?