I had a multifamily deal for a PE client that required building a comprehensive institutional-grade real estate model. The deliverable included monthly cash flows, annual proforma, development budget, detailed assumption, and waterfall tab. This is the kind of model that typically takes 1-2 full days to build from scratch, involving a lot of repetitive setup work before you even get to the analytical heavy lifting.

Why AI?

I knew the entire model might be challenging to build completely with AI, but I saw a clear opportunity: most of the basic setup is grunt work with less analytical depth. Setting up monthly cash flow structures, linking assumptions across tabs, creating development budget line items. These tasks are time-consuming but formulaic, exactly where AI excels. I decided to push the boundaries and see how much of the initial framework I could automate.

My Starting Point

I've been using AI tools like Claude for different research purposes for several months, but this was my first attempt at building a financial model from scratch using AI. I had moderate experience with AI prompting but had never tackled something this structured and technical.

Tools Used

I used two primary tools for this project:

Claude Sonnet 4.5 (paid version) for developing and refining the prompt structure. I spent time working with Claude to break down the model-building process into discrete, manageable prompts.

Shortcut.ai (free version) for executing the prompts and generating the actual Excel output. Shortcut.ai takes natural language prompts and converts them into working spreadsheet models with formulas and formatting.

The Actual Prompts

I've shared the exact prompts I used here: (Link Here)

The final structure consisted of 8 discrete prompts targeting 4 different tabs. Each prompt was carefully crafted to be specific, actionable, and focused on a single output. I didn't rush into execution. Instead, I spent 30-40 minutes upfront with Claude just finalizing the prompt structure before rushing to generate financial model with vague prompt.

The Process

Here's my step-by-step approach that made this work:

Data Inputs

I fed selective pages from the client's investment deck into the prompts. This included key financial assumptions, property details, and market data. I didn't dump the entire deck. Instead, I extracted only the relevant information needed for each specific prompt.

What Worked Well

There were some issues, though none were showstoppers:

Assumption errors: Despite having detailed information in the deck, the AI made mistakes on a few assumptions. However, these were easy to spot during my review, and fixing them was just a matter of changing a few cells.

Annual pro forma formulas: The Annual Pro Forma tab had a few formula issues as well. These were simple errors, easy to identify, and quick to fix once reviewed.

Assumptions tab layout: The assumptions section typically consists of 5–10 small tables arranged in a clean, dashboard-style layout. The AI generated each table sequentially, one below the other, so we had to reorganize the layout to make it presentation-ready.

Key metrics formulas: The formulas for calculating levered cash flow and unlevered cash flow were slightly off. They weren't aligned with how we calculate these metrics internally, so I had to manually correct the formulas.

Complexity limitations: I didn't even attempt the waterfall and sensitivity tabs with AI. These sections are inherently complex, with multiple return hurdles and intricate calculation logic. I used our existing templates for these, which was the right call.

Quality Assessment

The output quality was genuinely impressive for the scope attempted. The tabs the AI generated would be acceptable deliverables with minimal touch-up. The formulas were more sophisticated than I would have written myself. The structure followed institutional modeling standards without me having to specify every formatting detail.

That said, this wasn't a plug-and-play solution. The model still required analyst oversight, formula verification, and assumption validation. But as a foundation? It was excellent.

The Numbers

I completed the four tabs in a few hours compared to the typical 1-2 full days this would take building from scratch. While I don't have an exact hour count, the efficiency gain was substantial. The time saved allowed me to focus more on the analytical components, reviewing assumptions, and building out the more complex waterfall logic.

Key Learning

If I were to do this again, here's what I'd do differently:

Define formulas explicitly: For metrics like levered and unlevered cash flows, I would provide the exact formulas I want rather than assuming the AI would match our internal methodology.

Be more precise about assumptions: I would specify exact values and sources for key assumptions upfront rather than relying on the AI to interpret the deck.

When to Use vs. Not Use

Use this approach for: Initial model setup, repetitive structure building, standard cash flow frameworks, and any tab where the logic is formulaic rather than highly custom.

Don't use this approach for: Complex waterfalls with multiple return tiers, highly customized sensitivity analyses, or any modeling component where your firm has very specific calculation methodologies that differ from market standards.

Skill Requirements

To use this workflow effectively, you need:

Strong financial modeling fundamentals: You must be able to validate the AI's output. If you don't understand how a multifamily model should be structured, you won't be able to catch the AI's mistakes.

Ability to break down complex tasks: Success depends on decomposing your model into specific, manageable prompts. If you can't articulate what you need in discrete steps, the AI will struggle.

Attention to detail: You still need to review every formula, validate every assumption, and check every linkage. This is not autopilot.

Warnings and Pitfalls

Don't expect end-to-end output. The AI will get you 80-85% of the way there, but you can't skip analyst responsibilities. You're still accountable for accuracy and completeness.

Validate everything. AI can and will make mistakes. Some will be obvious, others subtle. Don't trust formulas blindly. Click through and verify logic.

Know your firm's standards. The AI generates 'market standard' models. If your firm has specific conventions or calculation methodologies, you need to explicitly specify these or be prepared to edit afterward.

Replicable Template

Yes, this workflow is absolutely replicable. The exact prompts I used are available in the link above. You can adapt them for your specific deal by swapping in your property details, assumptions, and investment structure.

Customization Tips

To customize this for different deals:

Start with the prompt structure, not the content. The framework of breaking the model into assumptions, monthly flows, annual flows, and development budget works for most multifamily deals.

Modify the data inputs section to reflect your specific deal parameters.

Add or remove tabs based on your deliverable requirements.

Tool Recommendations

I used the free version of Shortcut.ai, which was sufficient for this project. For larger or more complex models, the paid version might be necessary for additional features and capacity. Claude Sonnet 4.5 is available through the standard Claude.ai subscription.

How I Verified

Given this was my first serious attempt at AI-driven financial modeling, I was thorough with validation:

Formula audits: I clicked through every significant formula to verify the logic. I checked that cells referenced the correct inputs, calculations flowed in the right direction, and the mathematical operations matched what I intended.

Assumption verification: I cross-checked every assumption against the source deck. Any discrepancy was flagged and corrected.

Output sense checks: I ran basic reasonability tests. Do the cash flows trend correctly? Are the metrics in expected ranges? Do the development budget totals match the source materials?

Comparison Benchmarks

I compared the AI-generated tabs against models I've built manually for similar deals. The structure and approach were consistent. In some ways, the AI output was cleaner and more sophisticated than my typical manual builds. However, the manual models will suffice our purpose with high accuracy.

Confidence Level

For basic modelling tasks with straightforward assumptions and standard structures, I have high confidence in this approach. It's a genuine time-saver that produces quality output.

For complex scenarios with custom calculation methodologies or intricate waterfall logic, my confidence is lower. I would still use AI for the initial framework, but I'd plan for significant manual refinement and validation. The more your modelling deviates from market standards, the more hands-on work you'll need to do after the AI generates its output.

Would I trust this for real investment decisions? Yes, but only after thorough validation. The AI is a powerful accelerator, not a replacement for analytical rigor. Use it to handle the grunt work but bring your full expertise to verifying and refining the output.