Skip to main content

The grounded workflow

Claude for commercial real estate underwriting

Underwriting is where an AI workflow either earns a firm's trust or loses it. The setup PSV builds on is deliberately unglamorous: your deal documents, your template, and a source behind every figure that a person can check before committee sees it.

Get your free AI assessmentWhat is CRE AI?
Zed Truong

By , Co-founder, Pacific Software VenturesPublished · Updated

Short answer

Grounded in the documents, checked by a person.

Yes, Claude by Anthropic can support commercial real estate underwriting, and it is at its best when grounded in the actual deal documents: the OM, the T-12, and the rent roll. The approach PSV builds on is simple: Claude reads your files, fills your underwriting template, and cites the page or line behind every number, and a person verifies the work before it reaches committee. Ungrounded, no general model should be trusted with committee-ready figures. Grounded and cited, Claude is a strong first pass on the documents, with every number traceable to its source.

The workflow

Underwrite a deal with Claude, step by step

Four moves. None of them is a prompt trick, and all four are things your team controls.

  1. 01

    Give Claude the deal documents

    Upload the actual OM, T-12, and rent roll. A broker summary is not the deal, and an answer about a deal nobody has read is worth what it cost.

  2. 02

    State your rules and your template

    Spell out your underwriting assumptions and thresholds, and name the exact template to fill. The output arrives in your format, not the model's.

  3. 03

    Require a citation per figure

    Every rent, expense, and assumption gets the page or line it came from. A number without a source is a number you cannot use.

  4. 04

    Verify before committee

    A person checks the citations, chases the anomalies, and signs off. Claude drafts the first pass. The underwriter owns the numbers.

This is how PSV builds CRE AI on Claude by Anthropic: cited truth as a system property, in a workflow your team owns. The same setup is taught in CRE AI training and deployed for firms.

Division of labor

What Claude handles well, and what stays yours

Well-suited to Claude

  • Long documents: a full deal package read in one pass, first page to last
  • Rule-following: detailed, multi-step instructions held across the whole job
  • Abstraction: a messy rent roll turned into your clean, consistent template

Stays human-owned

  • The go or no-go decision
  • The assumptions behind the model
  • The negotiation

Weighing ChatGPT for real estate underwriting instead? The same grounding rules apply to any model. See the full Claude vs ChatGPT for CRE comparison.

Failure modes

Where AI underwriting goes wrong

Ungrounded answers

Asked about a deal nobody uploaded, a model will still produce a confident answer, and it will read like the real thing. The fix is structural, not a better prompt: an answer the deal documents do not support is not an answer.

Unverifiable numbers

A figure with no trail cannot be checked, so it cannot be trusted. A required citation per figure makes every number traceable to its source document.

Template drift

Left loose, outputs arrive in a different shape every time. A stated template and standing rules keep the work consistent, and keep the workflow yours.

All three trace to the same root: an ungrounded model with no obligation to show its work. The grounded setup removes the root, which is why it comes first in CRE AI done properly.

Why the Company Brain matters here

Grounding makes the numbers checkable. It does not make them yours.

A cited first pass on a deal package is a real gain, and it is still a stranger's underwriting. What separates your read on a deal from anyone else's is narrower than it sounds: which comps count as comparable here, and which assumptions your firm stands behind. The Company Brain is what makes that judgment ordinary, deal after deal.

  • A comp set your analysts would defend.

    The submarket lines your firm actually draws, the vintage and unit mix it will accept, and the buildings your team argued about last quarter and set aside. What comes back reads like work from someone who has been at the firm a while.

  • House assumptions you stop restating on every deal.

    Exit cap spread, reserves per unit, credit loss, the rent growth curve your committee will actually sign. Nobody retypes the same eight numbers to start a deal, and the argument in the room is about this asset instead of the defaults.

  • The comp set does not change with who pulled it.

    A first-year and a partner working the same submarket land on the same comparables and the same adjustments. When the house view on a submarket moves, the next deal reflects the new one instead of last year's.

See how the Company Brain works

The Company Brain, for underwriting

The model is yours. So are the assumptions inside it.

Underwriting is where a generic tool fails most visibly, because the answer depends entirely on assumptions your firm has spent years setting. The Company Brain holds them.

01

What your firm knows

The rent rolls, T-12s, and offering memoranda you have already worked through, the comps behind each one, and how the deal actually performed against the underwriting.

02

How your firm works

Your own model with its tabs and formulas intact, your standing assumptions, your return thresholds, and the sensitivity cases your investment committee expects to see.

03

How your firm decides

The adjustments you always make to a broker's numbers, the line items you never take at face value, and the reasoning that has to survive the committee.

Surfaces in

ClaudeChatGPTMicrosoft Copilot

Without it, you get a populated template built on someone else's assumptions, and every figure has to be re-checked before it can be trusted.

How the Company Brain works

Frequently asked

Claude for underwriting, answered

Can Claude underwrite a commercial real estate deal?
Claude can support underwriting: reading the OM, T-12, and rent roll, filling an underwriting template, and citing the source behind each figure. It should not replace the underwriter. A person verifies the citations and owns the go or no-go decision.
How do I underwrite a deal with Claude step by step?
Upload the actual deal documents, state your underwriting rules and the exact template to fill, and require a source citation for every figure. Then verify the citations before the numbers reach committee. The setup, not the prompt wording, is what makes the output dependable.
What documents does Claude need to underwrite a deal?
At minimum the offering memorandum, the trailing 12-month financials, and the current rent roll. Loan terms, comps, and market reports help when the analysis calls for them. Actual files beat summaries, because the model can only cite what it can read.
How accurate is Claude for real estate underwriting?
Accuracy depends more on the setup than the model. Grounded in the real documents and required to cite a source per figure, every number is checkable, and a person checks it. No general model's unverified output should be treated as committee-ready.
Is ChatGPT or Claude better for real estate underwriting?
PSV builds on Claude for long, rule-bound deal documents, where steady instruction following across a full document set matters most. Either model improves sharply with grounding, stated rules, and required citations. The full comparison is on the Claude vs ChatGPT for CRE page.
Does AI underwriting replace analysts?
No. The model handles the document-heavy first pass: extraction, abstraction into the template, and flagging what does not reconcile. Analysts own the assumptions, the verification, and the recommendation, which is where underwriting judgment lives.

Practice it

Learn this workflow on a real deal

The CRE AI Institute, by PSV, teaches this exact underwriting workflow on a real practice deal, with Claude reading the documents and citing its sources. 7-day free trial. $0 today.

Explore the InstituteWhat is CRE AI?

Keep reading

The rest of the CRE AI cluster

PSV Research Desk

Related PSV analysis

Source-cited reporting on the workflows in this guide.

UNDERWRITING

One Claude Prompt, Full Underwriting: What Has to Be True First

Can Claude or ChatGPT run a full underwriting workflow, intake to IC draft, from a single prompt? It is the single most common question in PSV’s prospect calls and Institute office hours, usually asked by a team that has seen each step work in isolation and wants the button that runs them all. The honest answer is yes, the chaining is the easy part, and the reasons it fails have nothing to do with the prompt.

AUG 16, 2026 / 8 MIN READ

Read

AI TOOL SELECTION

Claude vs Copilot: What CRE Firms Get Wrong About Both

Every CRE firm on Microsoft 365 is having the same meeting: IT wants Copilot because it is already in the tenant, an analyst wants Claude because of what they saw it do with a rent roll, and someone is asking why the firm should pay for both. The question gets argued as a rivalry. It is actually a category error, and the firms that see through it end up with a cheaper, better stack than the ones that pick a side.

AUG 16, 2026 / 12 MIN READ

Read

Deal desk notes · Weekly

Written for the people who own the numbers.

One CRE workflow taken apart each week: what to hand the model, what to require back, and where an analyst still has to check the work. Free, and one click to unsubscribe.

No spam. Unsubscribe in one click.

Next step

See what this would change in your underwriting.

Ten minutes. An honest read on where you stand, the highest-value opportunities, and which engagement fits, if any.

Get your free AI assessmentTalk to PSV

We use cookies for analytics. Your data helps us improve the site.