How Does AI Underwriting Work? Connect Documents, Evidence, and Excel

When I look at an AI underwriting system, I start with a simple question: can we follow a number from the final report back through the model to the evidence that informed it? A fluent explanation is useful only if that chain holds together. Our work at Cactus is largely about building and checking that chain.

This guide follows the multifamily broker underwriting workflow. It turns uploaded deal documents into structured evidence, combines that evidence with market context, maps chosen inputs to a configured Excel model, and prepares a report from the calculated results. Other asset classes and model configurations can have different requirements.

AI-powered commercial real estate underwriting requires reasoning across financial data, rather than treating every extracted number as a model input. An auditable AI financial modeling platform needs to show which source supports an assumption, why a conflicting value was not adopted, and how the chosen input reached the calculated result. That is the purpose of the evidence architecture described here.

What happens between an uploaded document and a report?

The boundary between stages matters. A parser can successfully read a PDF without proving that the financial statement covers the right period. An agent can choose a plausible assumption without proving that the workbook calculated successfully. Each stage has a different job.

What does the knowledge graph connect?

Think of the graph as a connected deal file. Shared documents sit under the deal. An individual underwriting has its own property context, model information, market context, and results. Those relationships help an agent ask a precise question, such as which rent roll supports this unit count, instead of searching an undifferentiated pile of text.

Before the primary underwriting analysis, the pipeline freezes references to versions of the property evidence, market evidence, and model contract. We call that collection the Deal Room. It provides a stable context for the run: later stages can read the evidence selected for that analysis instead of silently following a newer version.

This also keeps a calculated result distinct from an uploaded fact. A seller’s reported NOI and the model’s normalized NOI can both be useful. They describe different things, so storing one over the other would erase the explanation the buyer needs.

How do source facts stay separate from assumptions?

Extraction identifies what a document says. Underwriting decides how to use it. The agent instructions require attention to source authority, period, scope, and units. A T-12 expense total, a monthly ledger subtotal, and a broker’s forward projection should not be blended just because they share the label “expenses.”

Prepared rent-roll and financial rows still need source review. The workflow checks document references and whether the agent read the source it cites. Tests cover wrong document references, skipped rent rolls, duplicate units, and missing numeric values. A missing amount must not quietly become zero, and a zero-rent unit must not automatically become a vacant unit.

Defaults and market estimates are assumptions, with their limitations made explicit. When a supported fact conflicts with another source, the agent must explain its selection. When the conflict remains unresolved, the uncertainty needs to survive into the analysis.

Follow a 100-unit property through the system

Suppose an OM describes 100 apartments, but the supplied rent roll contains 98 distinct unit rows. The first task is to preserve both observations. The system should not invent two rows to make the rent roll match the brochure. The analyst needs to determine whether units are excluded, the files cover different dates, or the property description is wrong.

Next, suppose a complete T-12 supports $1.20 million of annual income and $450,000 of operating expenses. Those source totals imply $750,000 of reported NOI. If an underwriting adjustment raises the expense basis by $25,000, the model’s normalized NOI would be $725,000. The adjustment is a separate decision with a rationale; it does not rewrite the T-12.

The selected facts and assumptions map into the workbook’s declared input fields and tables. The Excel calculation engine evaluates the workbook formulas and returns defined outputs. For a simple direct-capitalization illustration, $725,000 divided by a selected 5% cap rate is $14.5 million. The 5% remains an underwriting assumption that needs support and sensitivity review.

Agents then use the returned results to explain the income, expense adjustments, valuation, and risks. They do not get to substitute a more attractive number in the narrative. Delivery checks reconcile reported income less expenses to reported NOI, the normalization bridge to normalized NOI, and the selected valuation basis to the result.

Which gaps need a disclosure, and which stop the run?

A missing document and an unreadable supplied document are different conditions. The pipeline can work with an incomplete package and disclose its limitations. But if a supplied document fails to read, the document-read gate can stop the run so the analysis does not proceed as if that document had been considered.

Business uncertainty can remain visible in a preliminary or qualified analysis. A failed calculation, incorrect workbook identity, or broken financial reconciliation requires a different response. The workflow includes deterministic audit checks and a bounded opportunity for the analyst to resolve issues before the publication decision.

What should a buyer review in the finished analysis?

• Check that reported and normalized income, expenses, and NOI use clearly identified bases.
• Inspect the assumptions and their support, including defaults, estimates, confidence, and sensitivity.
• Look for missing evidence and source-age notices before treating a result as decision-ready.
• Compare the report with the corresponding Excel output, and ask how a new document creates a new analysis version.

A knowledge graph is useful when it improves these review steps. The goal is an analysis whose evidence, assumptions, formulas, and explanation can be inspected together. Badreddine’s guide explains the Excel model integration that makes the calculation boundary work.