Claude in Excel Review

I'm in multifamily development and have been testing the Claude Excel plugin over the past few days. People were asking about it in a couple different discussions on the Off Topic board.

It requires a Pro subscription at $20/month. I was already paying for it since I use Claude regularly for memo writing and text analysis, so the Excel plugin was just a nice bonus for me.

The UI is clean and intuitive. Plugin runs on Opus 4.5, which is their most powerful model. One significant constraint - daily usage limits. Both days I tested it, I hit my limit within a couple hours of work.

I first tested it on our development model. Our development model is fairly complex. It has 15-20 tabs covering budget, cap table, operating assumptions, pro forma, waterfall, executive summary, etc. It's fully dynamic with automated processes and formatted outputs. Before testing, I spent 2-3 hours writing a comprehensive guide explaining the model structure and tab logic, then uploaded it as a "Skill" for Claude to reference. This definitively helps - it navigates the model faster when it has that context.

I gave it qualitative inputs for a 300-unit garden development (land cost, target unit mix solving for specific average SF, untrended rents, etc.) and asked it to build out the bones of the deal in the model. Results were mixed. It handled basic inputs well but struggled with the more complex request - creating a unit mix (in the model's unit mix section) that solves for a specific average SF. Got close but not exact. More concerning, it deleted a formula in one row while making changes, despite the Skill guide explicitly instructing it to never touch formula or link cells. That absolutely has to be fixed. 

Once the deal was built out, I asked it to build an exit cap rate sensitivity table showing how multiple outputs change across different cap rates. This didn't go well. It got stuck trying to create an Excel data table, which doesn't work efficiently in our model due to the number of tabs and waterfall calculations involved. The better approach is manually changing the cap rate and pasting values, rinse, and repeat. Even after I explained this method, it kept trying to force the data table solution.

The last thing I had it do was testing it on some CoStar data analysis. Our markets have heavy supply. Anecdotally many 2023-2024 lease-ups are stalling out at 80-90% occupancy and this is big issue in the market. I pulled historical data segmented by vintage (2016, 2017, etc.) and asked Claude to chart occupancy trajectories at +1 year and +2 years post-delivery to compare against recent vintages. It handled the +1 year chart fine, but when it tried to create the +2 year chart, it just recreated the same thing.

Despite the fact that it didn't work so well when I tested it, I'm actually cautiously optimistic for three reasons:

1) It's already better than Copilot, which is effectively useless for this type of work.

2) Claude Code had a similar trajectory - mediocre at beta version, now reportedly excellent. Also, the foundation model will only get stronger with Opus 5+.

3) Most importantly, while writing the model guide, I'd run drafts through Opus asking "does this make sense?" Its comprehension of complex formulas was impressive. It understood not just what formulas were doing, but why specific subformulas were necessary without me even mentioning the subformulas most of the time. That level of understanding suggests the limitation is fine-tuning in Excel not underlying brainpower.

TLDR - Still very much in beta. Not ready for production use on complex models, but not totally useless if you manage expectations and supervise carefully. Better than the competition. Anthropic's track record with improving Claude Code gives me confidence they'll iterate this into something useful. The underlying model is already powerful. It's just a matter of bringing that capability into Excel.