How to Get AI to Understand Your Deal Structure

AI understand deal structure when you explicitly describe your capital stack, return priorities, cash flow timing, and distribution logic in structured, unambiguous terms. Most AI errors in real estate models stem not from weak algorithms but from vague inputs—specifying "3 debt tranches with a 70% LTV" without defining seniority, payment schedules, or prepayment penalties leaves gaps the AI fills incorrectly.

Relevant Articles

• Need prompt templates? See How to Use Natural Language to Build Excel Pro Formas.
• Building a waterfall? Review How to Build a Waterfall Model Using AI.
• Want complete model output? See How to Get AI to Output Complete Models, Not Fragments.

Working Example: Deal "Avalon Ridge"

To see this in action, let's model a specific deal:

This deal is moderately complex. It has two debt tranches with different terms, a promote structure with two hurdles, and a planned refinance that affects the debt stack. If you ask AI to "model this deal" without breaking down each component, you will get generic formulas that ignore the mezzanine debt payoff timing or misallocate refinance proceeds.

Communicating Deal Terms

Start with the capital structure in priority order. Do not assume AI infers seniority from context. Many models fail because the prompt lists "senior debt" and "mezzanine debt" without stating that senior debt gets paid first on exit or that mezzanine interest accrues if cash flow is insufficient to cover current pay.

For Avalon Ridge, the correct specification looks like this:

Capital Stack (in order of payment priority):

1. Senior Debt: $24,000,000 at 4.5% fixed annual interest. Interest-only payments in Years 1-2. Amortizing payments Years 3-7 based on a 30-year schedule. Prepayable without penalty. Outstanding balance due at exit.
2. Mezzanine Debt: $2,500,000 at 9% annual interest, current pay (not accruing). Maturity in Year 5—if not refinanced, full principal plus accrued interest is due. If property refinances in Year 4, mezzanine debt is paid off in full from refinance proceeds before any equity distribution.
3. Equity: $12,000,000 total. LP contributes $10,800,000 (90%), GP contributes $1,200,000 (10%). Cash flow after debt service distributes per the waterfall (described below).

This structure explicitly states payment timing, accrual behavior, and event-driven payoffs. Without this level of detail, AI will default to simplistic assumptions—often treating all debt as a single blended rate or ignoring the mezzanine maturity constraint.

Common errors we see in 70% of prompts: listing debt amounts without defining payment schedules, omitting prepayment terms, and failing to specify what happens to mezzanine debt on refinance. In Avalon Ridge, the refinance in Year 4 triggers a full mezzanine payoff. If you do not state this, AI will continue accruing mezzanine interest through exit, inflating debt service and understating equity returns.

Specifying Capital Stack

The capital stack is not just a list of numbers. It is a decision tree. At every cash flow event—monthly operations, annual distributions, refinance, exit—the model must know who gets paid, how much, and in what sequence.

For Avalon Ridge, the capital stack logic translates to these calculation rules:

Monthly Operating Cash Flow Allocation (Years 1-7):

1. Pay senior debt service (I/O or amortizing per schedule).
2. Pay mezzanine debt interest ($2,500,000 × 9% / 12 = $18,750/month).
3. Distribute remaining cash to equity per waterfall.

Refinance Event (Year 4):

1. New senior debt: $28,875,000 (75% LTV based on appraised value of $38,500,000).
2. Pay off existing senior debt balance (approximately $22,100,000 after 4 years of partial amortization).
3. Pay off mezzanine debt in full: $2,500,000.
4. Remaining proceeds ($28,875,000 - $22,100,000 - $2,500,000 = $4,275,000) distribute to equity per waterfall.

Exit Cash Flow (Year 7):

1. Sale proceeds = Stabilized NOI / 5.75% cap rate.
2. Pay off senior debt balance (new loan from Year 4, with ~3 years of amortization).
3. Distribute remaining proceeds to equity per waterfall.

Notice the specification includes intermediate calculations. "75% LTV" is not enough—you must define LTV based on what value (purchase price, appraised value, or Year 4 stabilized NOI / cap rate). In institutional models, this is typically appraised value, but if you do not state it, AI may use original purchase price, yielding a materially different refinance amount.

The specification meta-skill applies here: defining "refinance at 75% LTV" is not specific. Defining "refinance at 75% of appraised value, where appraised value equals stabilized NOI divided by current market cap rate of 5.9%" is specific. The second version produces a defensible model. The first produces a guess.

Describing Cash Flow Logic

Cash flow logic is where most deal structures break. The issue is not complexity but ambiguity. A waterfall with "8% preferred return" sounds simple until you ask: is the pref calculated on contributed capital, unreturned capital, or cumulative contributions net of distributions? For Avalon Ridge, we use unreturned capital—the industry standard for institutional deals.

Here is the explicit logic for Avalon Ridge's equity waterfall:

Tier 1: Return of Capital + Preferred Return (8% on unreturned LP capital)

• LP receives 100% of distributions until:
  • LP has received $10,800,000 (return of contributed capital), AND
  • LP has received cumulative distributions equal to an 8% annual return on unreturned LP capital.
• Unreturned capital in any year = $10,800,000 - cumulative distributions to LP as of the prior year.
• Preferred return is calculated annually: Unreturned Capital × 8%. If Year 1 distributes $500,000 to LP, Year 2's pref is calculated on ($10,800,000 - $500,000) = $10,300,000 unreturned capital.

Tier 2: Split to 15% LP IRR (70% LP / 30% GP)

• After LP pref is satisfied, distributions split 70% LP / 30% GP until the LP achieves a 15% IRR on their $10,800,000 investment (measured from contribution date to each distribution date).

Tier 3: Split Above 15% LP IRR (60% LP / 40% GP)

• All remaining distributions split 60% LP / 40% GP.

This description is unambiguous. It defines the pref calculation method, specifies cumulative vs. annual hurdles, and ties IRR hurdles to specific cash flows. Without this, AI will default to simpler (and wrong) logic—often calculating pref on original capital instead of unreturned capital, which overstates LP returns.

In our models, we always separate the waterfall logic into three distinct Excel ranges: (1) a Returns block that calculates LP IRR and cumulative pref, (2) a Tier Test block that determines which tier applies in each period, and (3) a Distribution block that allocates cash per the tier result. This decomposition prevents circular references and makes verification straightforward. When specifying your deal structure to AI, describe this separation explicitly. Do not assume AI will infer the correct architecture.

Handling Non-Standard Structures

Avalon Ridge includes a non-standard feature: a refinance event in Year 4 that pays off mezzanine debt and distributes excess proceeds to equity. Many AI tools struggle with event-driven cash flows because they expect linear, annual structures.

To handle this, break the timeline into discrete cash flow events:

Event 1: Acquisition (Year 0)

• Equity contribution: $12,000,000 (LP: $10,800,000, GP: $1,200,000)
• Senior debt draw: $24,000,000
• Mezzanine debt draw: $2,500,000
• Total sources: $38,500,000 (matches purchase price)

Event 2: Operations (Years 1-4)

• Monthly NOI after debt service distributes to equity per waterfall
• Track cumulative distributions to calculate unreturned capital and pref accrual

Event 3: Refinance (Year 4, end of year)

• Trigger: Property appraised at stabilized NOI / 5.9% cap rate
• New senior debt: 75% of appraised value
• Use proceeds to: (1) pay off old senior debt, (2) pay off mezzanine debt, (3) distribute remainder to equity
• Post-refinance debt service adjusts to new senior debt terms (assume same 4.5% rate, fresh 30-year amortization)

Event 4: Operations (Years 5-7)

• Monthly NOI after senior-only debt service (no more mezzanine) distributes to equity
• Continue tracking cumulative distributions for waterfall tier tests

Event 5: Exit (Year 7)

• Sale proceeds = stabilized NOI / 5.75% exit cap
• Pay off senior debt balance
• Distribute remainder to equity per waterfall

This event-based structure forces you to define cash flow logic at each decision point. For the refinance, you must specify appraisal methodology, debt sizing rules, payoff priorities, and post-refinance debt terms. Each specification reduces ambiguity.

A common error in non-standard structures: assuming AI will "figure out" that excess refinance proceeds go to equity. If you do not explicitly state "after paying off all debt, remaining proceeds distribute per the equity waterfall," AI may stop at debt payoff and ignore the distribution, creating a phantom cash reserve. Always close the loop on every dollar.

Iterating to Accuracy

Your first specification will be incomplete. Accept this. The goal is not to write a perfect prompt on attempt one—the goal is to create a falsifiable structure you can test and refine.

For Avalon Ridge, here is the iteration process we used:

Iteration 1: Initial prompt included capital stack and waterfall structure but did not specify mezzanine payoff on refinance. AI model continued mezzanine debt through exit. Error identified during manual review of Year 4 cash flows.

Iteration 2: Added explicit instruction: "In Year 4 refinance, pay off mezzanine debt in full before distributing excess proceeds to equity." AI correctly paid off mezzanine but calculated excess proceeds as (new debt - old debt - mezzanine), omitting the senior debt balance reduction. Error identified by comparing refinance proceeds distribution to expected equity return.

Iteration 3: Clarified refinance waterfall: "Refinance proceeds first pay off existing senior debt balance (calculate balance after 4 years of amortization), then pay off mezzanine debt in full ($2,500,000), then distribute remaining amount to equity per waterfall." Model now matches expected cash flows.

Each iteration added one layer of specification. This is decomposition in action: instead of trying to describe the entire refinance logic in one sentence, break it into sequential steps. AI executes sequential logic more reliably than nested conditionals.

The verification meta-skill is critical here. After each iteration, run a returns summary to confirm LP IRR and GP promote align with your expectations. For Avalon Ridge, we expected LP IRR around 16-18% based on underwriting assumptions. The first iteration returned 14% (because mezzanine debt overstated costs). The third iteration returned 17.2%—within tolerance.

Validating AI Understanding

Before you finalize the model, run these three validation tests:

Test 1: Capital Stack Reconciliation

At acquisition, total sources must equal total uses. For Avalon Ridge:

If this does not balance, your capital stack is incomplete or double-counted.

Test 2: Waterfall Tier Progression

Manually calculate LP IRR at the end of each year. Confirm the model assigns the correct tier:

• Year 1-3: LP has not yet received full return of capital + pref. Should be in Tier 1 (100% LP).
• Year 4 (post-refinance): Check if LP IRR exceeds 8% but is below 15%. Should be in Tier 2 (70/30 split).
• Exit (Year 7): If final LP IRR is 17.2%, final distribution should be in Tier 3 (60/40 split).

If distributions in Year 4 use a 60/40 split, the model skipped Tier 2. This indicates the hurdle test is misconfigured.

Test 3: Debt Payoff Timing

At exit, confirm all debt balances are zero:

• Senior debt balance after 7 years of amortization (starting Year 3 on original loan, then fresh amortization from Year 4 refinance) should be paid off from sale proceeds.
• Mezzanine debt should show zero balance from Year 4 onward (paid off at refinance).

If mezzanine debt shows a balance at exit, the refinance payoff logic failed. Go back to Iteration 2 above and re-specify.

These tests do not require complex formulas. They are logic checks. In institutional diligence, the first thing we verify is whether the model's internal cash flows reconcile at each event. A model that passes these three tests is structurally sound. A model that fails even one is wrong—not "close enough" or "directionally correct," but wrong.

This is the verification framework Apers teaches: build falsifiable tests before you review formulas. If the cash flows reconcile and the waterfall tiers progress logically, the formulas underneath are likely correct. If the tests fail, do not debug formulas—re-specify the structure.

Getting AI to understand your deal structure is not about teaching it real estate finance. It is about removing ambiguity from your description. Every term you leave undefined—"preferred return," "refinance," "promote"—is a gap AI will fill with a default assumption that may not match your deal. Specify priority, timing, calculation method, and event logic. Then verify that the output matches your specification. The tighter your input, the more accurate the model.