I tried building an Equity Waterfall model using AI recently and failed at it

This article explores a practical lesson from attempting to build a financial model using AI. While AI tools can accelerate financial modeling and create impressive structures, they still rely heavily on the clarity and completeness of user instructions. The piece draws parallels with onboarding junior analysts, highlights where AI adds value (formatting, structure), where it falls short (judgment, assumptions), and why domain expertise remains critical—especially in complex U.S. commercial real estate (CRE) modeling.

When AI Follows Instructions Too Well

I recently tried building an equity waterfall model using Perplexity Computer. The result? It didn’t match what I had in mind.

At first glance, it felt like a failure. But on closer inspection, the AI had done exactly what I asked—it just didn’t do what I intended.

The gap wasn’t in execution. It was in instruction.

This is a familiar situation in teams. You brief a new analyst on building a model—say, a promote structure for a multifamily deal. You assume certain conventions are “obvious”:

• Preferred return compounding method

• Catch-up structure specifics

• IRR hurdles and timing nuances

• Treatment of fees, promote splits, and capital events

The analyst delivers a model that technically works—but doesn’t align with your expectations. Not because they lack skill, but because the brief was incomplete.

AI behaves the same way—except it executes with zero ambiguity tolerance.

The Hidden Complexity of CRE Equity Waterfalls

In U.S. CRE, equity waterfalls can vary significantly:

• A typical institutional deal might include a 8% preferred return, followed by a GP catch-up, then tiered splits like 70/30, 60/40, and 50/50.

• Timing conventions (monthly vs quarterly compounding) can materially impact IRRs.

• Return of capital sequencing and promote crystallization points can shift investor outcomes.

  
  

When prompting AI, missing even one of these assumptions can lead to structurally “correct” but economically inaccurate outputs.

For example: If you don’t explicitly define whether the preferred return is cumulative and compounding, the AI might assume a simple return—leading to a materially different distribution schedule.

Where AI Actually Shines

Despite the initial mismatch, the exercise revealed where AI adds real value:

• Model structuring and formatting: AI can quickly generate clean, presentation-ready templates that resemble institutional-grade models used by U.S. REPE firms.

• Speed of iteration: Instead of starting from scratch, you get a working skeleton that can be refined.

• Idea scaffolding: It helps translate conceptual deal structures into a tangible starting point.

  
  

This is particularly useful in early-stage underwriting or when prototyping deal structures—for instance, testing different promote tiers for an acquisition.

Where Human Expertise Still Matters

AI does not replace the need for a financial modeler—it amplifies the importance of one.

• Instruction design requires expertise: You need a deep understanding of CRE mechanics to even know what to specify.

• Validation is non-negotiable: Every formula, assumption, and linkage must be checked—especially in high-stakes deals.

• Judgment cannot be outsourced: AI cannot infer intent behind ambiguous instructions or challenge flawed assumptions.

  
  

In practice, this mirrors how senior professionals review models built by analysts. The tool has changed, but the oversight requirement has not.

The Real Takeaway

AI is not a shortcut to skipping fundamentals—it is a force multiplier for those who already understand them.

• Building financial models is faster and more efficient.

• But writing effective instruction is itself a specialized skill.

• And reviewing outputs remains a human responsibility.

  
  

The biggest shift is not in modeling—it’s in thinking. You are no longer just building models; you are designing instructions for machines to build them.

Final Thought

The next time AI gives you the “wrong” output, it’s worth asking: did it misunderstand—or did it understand exactly what you said?

Because in most cases, the answer will tell you more about your own assumptions than the tool itself.