UNDERWRITING

AI Underwriting for Commercial Real Estate: A Source-Cited Workflow

How to use AI to normalize a rent roll and T-12, prepare a first-pass underwriting, and accelerate IC work while keeping calculations, assumptions, and approvals reviewable.

BY PSV RESEARCH DESKEDITED BY ZED TRUONGJULY 12, 2026UPDATED JULY 28, 20266 MIN READ

Direct answer

Let AI extract, normalize, reconcile, and explain. Keep financial calculations deterministic and every investment assumption explicitly owned by a person.

The first job is normalization, not valuation

A CRE underwriting package is messy by default. Property names drift across files, the T-12 uses ownership-specific account names, the rent roll may mix monthly and annual fields, and the OM may present a stabilized case as if it were current. AI is valuable first as an abstraction and normalization layer.

Row of brick and stone apartment towers along the Manhattan bank of the East River on a clear day, with a slim modern high-rise at right, trees and a highway ramp below, and river water in the foreground. — IMAGE: JAKUB HAŁUN / CC BY 4.0Buildings like these reach an underwriter as a PDF offering memorandum and a rent roll. A source-cited workflow keeps every line of the pro forma pointing back to the page it came from. Image: Jakub Hałun / CC BY 4.0.

The system should inventory every source, identify its as-of date, preserve the reported value, and map it into an institutional category without overwriting the original. If a field is ambiguous, it should be flagged rather than inferred. That discipline creates an audit trail before a return calculation ever runs.

Rent roll: unit or suite, tenant, area, in-place rent, lease dates, concessions, and status.
T-12: reported line item, institutional category, reported amount, normalization, and source reference.
OM: property facts and seller claims stored separately from verified operating data.
Open items: every missing, conflicting, or low-confidence field visible in one queue.

Run the math outside the language model

A language model can explain a cap-rate bridge or draft a sensitivity narrative. It should not be trusted as the only calculator for a multi-step financial model. Once inputs are normalized, formulas should run in Excel, code, or another deterministic calculation layer where a reviewer can inspect them.

This separation is essential: AI handles unstructured documents and narrative reasoning; the model handles arithmetic; the underwriter owns assumptions. The output should show which values came from sources, which were calculated, and which were chosen by the investment team.

Comparison of how core commercial real estate underwriting jobs are handled by hand, by a general chatbot, and by a grounded AI workflow.
The job	Done by hand	General chatbot	Grounded AI workflow
Rent roll abstraction	Analyst retypes unit, tenant, area, rent, lease dates, and status; underwriter hours go to transcription.	Can read and summarize the file, but ambiguous fields get inferred and mixed monthly and annual rents slip in.	Extracts unit, tenant, area, in-place rent, dates, and status, preserving originals and flagging ambiguity.
T-12 normalization	Underwriter remaps ownership-specific account names into house categories line by line for each deal.	Can propose category mappings, but reported values risk being overwritten and no source reference is attached.	Maps each line to an institutional category, keeps the reported amount, and attaches a source reference.
Rent roll vs T-12 reconciliation	Reviewer cross-checks scheduled rent against the rental-income run rate manually, before any projection.	Can compare the two sources, but differences are not quantified or tied to the documents needed to close them.	Automated checks quantify each difference and propose the documents or questions required to close it.
Return calculations	Underwriter builds the model in Excel; formulas are deterministic but start from manually keyed inputs.	Can explain a cap-rate bridge, but should not be the only calculator for a multi-step financial model.	Normalized inputs feed Excel or code; reviewers inspect formulas and the underwriter owns assumptions.
Investment-committee package	Analyst compiles the memo and support by hand, trading judgment time for assembly and transcription work.	Can draft an IC narrative, but material statements are not traceable to a file, page, tab, or cell.	Produces a source index, assumption register, normalization log, open questions, and a draft IC narrative.

Reconcile before you project

The most useful automated checks happen before the five-year cash flow: rent-roll scheduled rent versus the T-12 rental-income run rate, occupied units versus physical occupancy, reported management fees versus the underwritten fee, and stated lease terms versus the abstracted leases.

An empty double-height boardroom in an office building: a long charcoal table ringed by taupe ribbed-back chairs with white leather seats, set beside a faceted floor-to-ceiling glass wall and a long dark padded bench that curves along the windows, with green glass panels filling the left wall and a switched-off flat screen on the far wall. — IMAGE: BENJAMIN CHILD / CC0This is the room the numbers have to survive. Deterministic calculations and named assumption owners keep the investment committee, not the model, in control of every line it is asked to approve. Image: Benjamin Child / CC0.

A reconciliation does not need to resolve every difference automatically. It needs to make the difference visible, quantify it, and propose the documents or questions required to close it. That is how AI reduces review time without hiding uncertainty.

Prepare the investment-committee trail

A first-pass underwriting should end with more than a model. It should produce a source index, assumption register, normalization log, open-question list, and draft IC narrative tied to the model outputs. Every material statement should be traceable to a file, page, tab, or cell.

The goal is not autonomous investment judgment. It is a faster path from raw deal package to a reviewable decision, with the human underwriter spending time on risk, market, basis, and business-plan judgment instead of transcription.

The production standard applied to first-pass underwriting

The six controls PSV applies to every production AI workflow land cleanly here, because a disciplined underwriting already produces most of the artifacts they require. The named job is a deal package in, rent roll, T-12, and OM, and a first-pass underwriting out, with the underwriter owning the next decision. The source boundary is the package itself: every file inventoried with its as-of date, seller claims from the OM kept apart from verified operating data, ambiguous fields flagged rather than inferred into the model.

The layered output is native here. Sourced facts live in the source index, each traceable to a file, page, tab, or cell. Deterministic calculations run in Excel or code, where a reviewer can inspect the formulas. Working assumptions sit in the assumption register, owned by the investment team, and the recommendation is the draft IC narrative. Separating the four layers lets a committee challenge the correct one: the number, the formula, the premise, or the judgment.

The exception path runs through the open-items queue. Scheduled rent that does not reconcile to the T-12 run rate, occupancy that disagrees across sources, a lease term that conflicts with the abstract: each becomes a quantified difference, paired with the document or question needed to close it, before the five-year cash flow runs. The approval record is the IC trail itself; the normalization log, assumption register, and open-question list travel with the model, so the committee inherits the evidence, not only the outputs.

Measurement follows decision quality, not draft polish: whether reconciliation caught differences early, how often reviewers made material corrections, and whether the package reached a defensible decision faster. The full statement of this standard is published separately on the PSV newsroom as "The PSV Production Standard: How AI Work Ships in Commercial Real Estate Operations." This guide applies it to one job; the controls transfer to the next.

The operator read

Finish with the judgment call.

The reported facts are above. Enter your email for PSV's operator conclusion and a year of access to every newsroom brief.

THIS ARTICLEALL PSV BRIEFSNO CARD REQUIRED

Send the operator read

Free. No card. Privacy policy. Unsubscribe anytime.

Dispatch the read

Keep the operator view in your deal room.

by PSVLive

Build the workflow behind this analysis. Train with the operators doing it.

Real deal files, live builds, and production-ready CRE AI workflows inside the CRE AI Institute.

Explore the Institute

Our members come from teams at

7-DAY FREE TRIAL, CANCEL ANYTIMEFOUNDING RATE $69/MO, LOCKED FOR LIFE30-DAY MONEY-BACK GUARANTEE

Clear answers

Common questions about AI underwriting commercial real estate

Can AI underwrite commercial real estate?

AI can inventory documents, extract and normalize fields, reconcile sources, flag exceptions, and draft a cited first-pass narrative. Deterministic formulas should calculate returns, while an underwriter explicitly owns market, financing, and business-plan assumptions.

Which documents should a CRE AI underwriting workflow use?

The minimum packet usually includes the rent roll, trailing operating statement, offering materials, property facts, and the firm's underwriting template. The workflow should record each file's date and treat seller claims separately from verified operating data.

How do you prevent errors in AI underwriting?

Preserve reported values, attach a source reference to material fields, run calculations outside the language model, reconcile the rent roll to operating statements, and maintain an open-question register. Ambiguous or conflicting inputs should be flagged instead of silently inferred.

Primary sources and operating references

These references support the control, research, and operating standards used in this guide. PSV’s workflow recommendations are original analysis.

Topics

UNDERWRITINGAI underwriting commercial real estateAI for commercial real estate underwritingcommercial real estate AI underwriting

DUE DILIGENCE

AI for Commercial Real Estate Due Diligence: Build the Exception Desk

A source-controlled approach to document inventory, lease and financial reconciliation, environmental and physical review, open-item tracking, and decision-ready diligence reporting.

Read AI for commercial real estate due diligence

CRE AI SOLUTIONS

What a Commercial Real Estate AI Solution Must Prove Before It Goes Live

The best CRE AI solution is not the one with the most impressive demo. It is the one that produces work a principal, investment committee, or client can inspect and trust.

Read commercial real estate AI solutions

ASSET MANAGEMENT

AI for Commercial Real Estate Asset Management: A Portfolio Operating System

A practical framework for variance review, lease-event monitoring, capital planning, property-manager follow-up, and investor reporting across a CRE portfolio.

Read AI for commercial real estate asset management

Cornerstone PSV guides on the workflows in this article.

Claude for Real Estate Underwriting

The grounded underwriting workflow: real deal documents in, your rules applied, a source citation behind every number, a person signs off.

Read the guide

What is CRE AI?

The definitional guide to commercial real estate AI: the core workflows, how it differs from a chatbot, and how to learn it.