I Let AI Build My Underwriting Model - Here’s What Happened
This article explores the real-world use of AI (specifically Claude Opus via Perplexity Computer) in building a multifamily underwriting model. It highlights where AI excels—speed, structure, and baseline accuracy—and where it struggles—iteration, refinement, and handling nuanced real estate logic. The piece outlines why domain expertise remains critical and how AI currently fits best as a starting tool rather than a full replacement.
The Promise: Institutional-Grade Models in Minutes
AI has quickly moved from novelty to practical tool in commercial real estate underwriting. In this case, a multifamily model built using Claude Opus demonstrated something striking: the ability to generate a clean, institutional-quality model almost instantly.
The output resembled what you would expect from a seasoned acquisitions analyst at a private equity real estate firm:
Clearly structured tabs (Assumptions, Rent Roll, Cash Flow, Returns)
Color-coded inputs (blue for user inputs, black for formulas—standard convention)
Well-organized tables with consistent formatting
Logical grouping of metrics such as IRR, equity multiple, DSCR, and yield on cost
For context, junior analysts at a private equity firm often spend weeks refining formatting standards. AI compressed that effort into minutes.
Even more interesting, the model included metrics that were not explicitly requested. For example:
Yr 1 Opex Composition
Economic Occupancy
Other Income as a % of EGI and more
This suggests that AI has absorbed patterns from institutional underwriting practices across the market.
Where AI Shines: Speed, Coverage, and Baseline Accuracy
Three clear strengths emerged from the exercise:
Professional Presentation
The model was visually polished and aligned with institutional standards. This is non-trivial - presentation matters in investment committees, and AI nailed it.
Comprehensive Metric Selection
AI anticipated what an investor would want to see. It could create a Dashboard having - Levered IRR, Equity multiple, Cash-on-cash return, DSCR trends, Exit cap sensitivity.
The model included many of these without explicit prompting.
Largely Accurate Math
Core calculations—NOI, cash flow, return metrics—were mostly correct. This is a major leap from earlier AI tools that frequently broke basic financial logic.
For a first draft, this represents a significant productivity gain. In practice, this could cut early-stage underwriting time by 60–80%.
Where It Breaks: Iteration Is the Real Bottleneck
The real challenge began when moving beyond the first draft.
Patchwork Instead of Refinement
When asked to adjust how concessions and loss-to-lease are handled using a detailed rent roll—a common requirement in multifamily deals—the model technically complied.
However, instead of integrating the logic cleanly:
New calculations were layered on top of old ones
Redundant fields appeared
Flow between rent roll and pro forma became inconsistent
In institutional underwriting, rent roll integration is critical. AI struggled to restructure the model holistically. It added features, but didn’t “re-architect” the system.
The Infinite Debugging Loop
A second issue emerged with dashboard charts that failed to populate correctly.
Attempts to fix them led to:
Repeated instructions
Partial fixes
New errors introduced elsewhere
This reflects a broader limitation: AI can generate code or formulas, but debugging interconnected spreadsheet logic—especially across multiple tabs—is still unreliable.
Key Takeaways for CRE Professionals
This experience highlights several practical lessons for real estate investors and analysts.
Prompt Quality Directly Impacts Output Quality
AI is only as good as the instructions it receives. Vague prompts lead to generic models.
For example, specifying: “Model concessions as a % of gross potential rent declining over 12 months” will produce far better results than: “Include concessions”
Domain Expertise Is Still Essential
Without a solid understanding of underwriting users may not catch incorrect assumptions and AI may introduce unintended logic
Always Audit the Math
Decisions worth millions depend on the model. Leaving everything to AI would not be wise. Auditing the model is essential.
The Right Way to Use AI in Multifamily Underwriting
The biggest insight is not that AI fails—but that it fits a specific role.
Today, AI is best used as:
A first draft generator
A structure and formatting assistant
A brainstorming partner for metrics and layout
It is not yet reliable as:
A final underwriting tool
A fully autonomous modeling system
A substitute for analyst judgment
Think of it like a highly efficient junior analyst who works fast but needs supervision and correction. It is not a tool for someone with no modelling background to build models.
Comments