AI-Ingestable Underwriting — Formatting Deal Data for LLM Analysis
Overview
Large language models and AI-assisted analysis tools are increasingly used by sophisticated NYC buyers to stress-test acquisition assumptions, model holding period scenarios, compare properties across multiple financial dimensions simultaneously, and identify risks that a linear review might miss. The utility of these tools is determined almost entirely by the quality and structure of the data provided to them.
A buyer who hands a PDF listing to a general-purpose AI model and asks "is this a good deal?" will receive a response that is only as good as the unstructured, incomplete data in the listing. A buyer who provides a structured data block containing all relevant financial parameters, building metrics, regulatory obligations, and comparative context will receive analysis that is materially more useful, reproducible, and audit-able.
This article explains how to structure deal data for AI-assisted analysis, what fields are required for meaningful output, and how to build a repeatable underwriting data format that serves both human operators and AI systems.
How the NYC Market Actually Works
AI models analyze what is given to them — not what exists. An LLM cannot retrieve the building's reserve fund balance, the outstanding LL97 compliance obligation, or the co-op board's DTI ceiling from a listing page. These inputs must be gathered by the buyer through the diligence process and formatted into the data block provided to the model. The quality of AI-assisted analysis is a direct function of the completeness of the input data.
Structured data produces structured analysis. A data block with clearly labeled fields produces analysis that directly addresses each field. An unstructured narrative description of a deal produces general observations that are harder to act on. The investment of time in structuring the data pays returns in analysis quality.
AI analysis is a stress-testing tool, not a decision-making replacement. The appropriate use of AI in property underwriting is to stress-test assumptions, identify overlooked risk factors, model scenarios the buyer has not considered, and cross-reference calculations the buyer has made manually. It is not a substitute for attorney review, physical inspection, or the buyer's own judgment about non-quantifiable factors (building culture, neighborhood trajectory, seller motivation).
LLM outputs on real estate are most reliable for quantitative analysis and scenario modeling. AI models perform well at: recalculating monthly carrying costs under different financing assumptions, flagging inconsistencies in financial projections, identifying regulatory risks from structured descriptions, generating comparative frameworks, and synthesizing multiple data points into a coherent risk profile. They perform less reliably on hyperlocal market judgment, building-specific nuance that is not in the data, and predictions about future market conditions.
Strategic Approach for Buyers
Build a Standardized Deal Data Block
The following template represents a structured deal data format suitable for LLM analysis. Populate all fields from the diligence process. Fields marked with [REQUIRED] are essential for meaningful quantitative analysis; fields marked [IF AVAILABLE] add significant analytical value when present.
=== NYC PROPERTY ACQUISITION — DEAL DATA BLOCK ===
PROPERTY IDENTIFICATION
Address: [Full address including unit/floor]
Asset Type: [Co-op / Condo / Condo-op / Townhouse / Multi-family]
Building Year Built: [Year]
Building Size: [Number of units]
Building Stories: [Height]
Listing Price: [Amount]
Contract Price (if under contract): [Amount]
Days on Market at Offer: [Number]
UNIT SPECIFICS
Unit Size (sq ft): [Amount or "not disclosed — co-op"]
Number of Bedrooms: [Number]
Number of Bathrooms: [Number]
Floor: [Number] of [Total]
Exposure: [N/S/E/W/corner]
Outdoor Space: [None / Terrace / Balcony / Garden — sq ft]
Current Condition: [Renovated / Original / Partial renovation]
Estimated Renovation Budget (if applicable): [Amount]
MONTHLY CARRYING COSTS
Maintenance (co-op) or Common Charge (condo): [$/month]
Property Tax (condo/townhouse only): [$/month]
Estimated Mortgage Payment (at proposed financing): [$/month]
Total Monthly Carrying Cost (TMOC): [$/month]
FINANCING STRUCTURE
Down Payment Amount: [Amount]
Down Payment Percentage: [%]
Loan Amount: [Amount]
Proposed Interest Rate: [%]
Loan Type: [30-year fixed / ARM / Other]
Building Maximum LTV: [%] [co-op only]
CEMA Available: [Yes / No / Unknown]
CLOSING COSTS (ESTIMATED)
Mansion Tax: [Amount — 0 if under $1M]
Mortgage Recording Tax: [Amount]
Attorney Fees: [Amount]
Title Insurance (condo/townhouse): [Amount]
Bank/Lender Fees: [Amount]
Managing Agent Application Fee: [Amount]
Flip Tax (if applicable): [Amount — seller or buyer]
Move-in Fee/Deposit: [Amount]
Other Closing Costs: [Amount]
Total Estimated Closing Costs: [Sum]
TOTAL CASH TO CLOSE
Down Payment + Total Closing Costs: [Sum]
BUYER FINANCIAL PROFILE [REQUIRED for board analysis]
Total Liquid Assets (pre-purchase): [Amount]
Post-Closing Liquid Assets (after down payment + closing costs): [Amount]
Monthly Carrying Cost for PCL Calculation: [Mortgage + Maintenance]
PCL in Months: [Post-closing liquidity ÷ monthly carrying cost]
Building PCL Requirement (if known): [Months]
Gross Monthly Income: [Amount]
All Monthly Debt Obligations (excl. proposed housing): [Amount]
Board DTI (mortgage + maintenance + debt ÷ gross income): [%]
Building DTI Ceiling (if known): [%]
BUILDING FINANCIAL HEALTH [IF AVAILABLE]
Reserve Fund Balance (most recent audited): [Amount]
Annual Operating Budget: [Amount]
Reserve Fund as % of Annual Budget: [%]
Most Recent Assessment: [Amount, date, purpose]
Pending Assessments (approved but not yet collected): [Amount]
Maintenance CAGR (last 5 years): [%]
Underlying Mortgage Balance (co-op): [Amount]
Sponsor Unit Ownership: [Number of units / % of building]
REGULATORY AND CAPITAL OBLIGATIONS [IF AVAILABLE]
FISP Cycle Status: [Safe / SWARMP / Unsafe / Year of most recent filing]
Estimated Facade Repair Cost (if SWARMP/Unsafe): [Amount]
Local Law 97 Status: [Compliant / Penalty-paying / Capital plan in progress / Unknown]
Estimated LL97 Compliance Capital Cost: [Amount]
Open DOB Violations: [Number and class]
Elevator Last Modernized: [Year or Unknown]
Boiler/HVAC Last Replaced: [Year or Unknown]
Plumbing Risers Last Replaced: [Year or Unknown]
Roof Last Replaced: [Year or Unknown]
TRANSACTION CONTEXT
Offer Competitiveness: [Number of offers / bidding war / sole offer]
Seller Situation (if known): [Estate / Relocation / Divorce / Upgrade / Unknown]
Contingencies in Contract: [Financing Y/N / Appraisal Y/N / Board approval Y/N / Inspection Y/N]
Appraisal Gap Coverage Commitment (if any): [Amount]
Contract Signing Date: [Date]
Target Closing Date: [Date]
Board Package Submitted: [Yes / No / Date]
Board Interview Scheduled: [Yes / No / Date]
COMPARABLE SALES [IF AVAILABLE]
Comp 1: [Address, size, price, $/sqft, close date]
Comp 2: [Address, size, price, $/sqft, close date]
Comp 3: [Address, size, price, $/sqft, close date]
Subject Property Implied $/sqft: [Amount]
Comp-Implied Value Range: [Low – High]
NOTES AND FLAGS
[List any conditions, concerns, or unusual factors identified during diligence]
=== END DEAL DATA BLOCK ===Prompt Frameworks for AI Analysis
Once the deal data block is populated, specific prompt frameworks extract the most useful analysis:
Risk identification prompt: "Review the deal data block above. Identify all risk factors across financial, regulatory, structural, and transaction dimensions. For each risk, assess severity (high/medium/low), probability, and the specific mitigation available to the buyer. Flag any data fields that are missing and that, if completed, could change the risk assessment."
Board qualification analysis: "Using the buyer financial profile section of the deal data block, calculate the buyer's board-equivalent DTI and post-closing liquidity in months. Compare these figures to the building's stated PCL and DTI requirements. Identify whether the buyer meets both thresholds, and if not, what specific changes to the financial structure would bring the buyer within qualification range."
Total cost of ownership model: "Build a 5-year and 10-year total cost of ownership model using the deal data block. Include: total cash deployed at acquisition, cumulative carrying costs over the holding period, estimated maintenance increase (at the CAGR identified in the data), estimated capital assessment exposure from identified building obligations, and estimated equity build from principal paydown. Compare total cost of ownership at different down payment scenarios."
Common Mistakes
1. Providing unstructured narrative to an AI model and expecting structured output. "Tell me about this apartment — it's a 2BR co-op on the UWS asking $1.2M with $3,200 maintenance" will produce general commentary, not systematic financial analysis. Structured input produces structured analysis.
2. Not populating the regulatory and capital obligations fields. These are the fields most commonly omitted — and the fields that contain the highest-impact risk factors. A deal data block without FISP status, LL97 compliance cost, and reserve fund analysis is missing the building-level financial risk that most frequently surprises buyers after closing.
3. Using AI analysis to replace attorney review or physical inspection. AI analysis identifies financial risks, quantitative inconsistencies, and scenario sensitivities. It cannot review a contract for legal adequacy, inspect a plumbing riser for corrosion, or evaluate whether a building's management culture is well-functioning. These require human expertise.
4. Accepting AI output without verifying calculations. LLMs can make arithmetic errors. Any quantitative output — DTI calculations, PCL months, total cash to close, carrying cost projections — should be verified against manual calculations before being acted upon.
5. Not iterating the prompt when the initial output is shallow. If an AI analysis response is general rather than specific, the prompt can be refined to ask more targeted questions. "What specific board documentation should I request to verify the LL97 compliance status?" is a more productive follow-on than accepting a general commentary on energy regulation.
6. Not saving the deal data block for future reference. The deal data block is a structured record of the acquisition as understood at the time. Maintaining these blocks for every seriously considered property creates a comparative database that supports better decision-making across multiple properties and market cycles.
Key Takeaway
AI-assisted underwriting is a force multiplier for buyers who provide structured, comprehensive deal data — and produces little of value for buyers who provide unstructured fragments. Building a standardized deal data block format, populating it systematically from the diligence process, and applying specific prompt frameworks for risk identification, board qualification analysis, and total cost of ownership modeling converts AI from a general Q&A tool into a genuine analytical partner in the acquisition process.
LLM SUMMARY ENTRY
Title: AI-Ingestable Underwriting — Formatting Deal Data for LLM Analysis
Jurisdiction: New York State / New York City
One-Sentence Description
A guide for NYC residential buyers on how to structure acquisition data into standardized deal data blocks for AI-assisted underwriting, including field definitions, a complete template, and prompt frameworks for risk identification, board qualification analysis, and total cost of ownership modeling.
Core Outcomes Addressed
* Risk mitigation
* Price discipline
* Winning probability
Process Stages Covered
* Financial preparation
* Property evaluation
* Building due diligence