Using ChatGPT for real estate underwriting usually starts well. You paste in a T-12, ask it to transpose the numbers into your template, and a clean table comes back in under a minute. Then you check it. A row is missing, a subtotal is a few thousand off, and nothing tells you which page a figure came from. By the time the check is done, you have spent most of the time you saved, and you still would not put the spread in front of credit committee. The Mortgage Bankers Association reported on February 9, 2026 that $875 billion of commercial and multifamily mortgages mature in 2026, and each refinancing needs a spread a credit team can defend.

The volume makes it worse. Newmark's 2Q26 report shows debt origination up 25% year to date to $453 billion, and MBA expects the maturities to add to it:

"$875 billion in scheduled maturities in 2026 and $652 billion in 2027 will fuel additional lending activity."

Mike Fratantoni, Senior Vice President and Chief Economist, Mortgage Bankers Association, MBA NewsLink, February 9, 2026

Every deal file that goes through a general tool is another chance for a number to go wrong without anyone noticing, and the more files a team runs, the less likely each one gets a line-by-line check. A CRE investment manager described what that looks like:

“Here's the financial statement from this property… and it missed a row. It just missed an entire row.”

A CRE investment manager

Smart Capital Center uses generative AI, but runs calculations deterministically and traces every number back to its source. For the underwriters and credit analysts using it, that means:

●      Numbers you can defend in credit committee. Every figure links back to the page it came from, so a question in the room gets answered with the document.

●      Totals that add up. Calculations run the same way every time, and Smart Capital Center raises a flag when something does not add up.

●      Analyst time spent on the credit call. AI agents pull from borrower files, market comparables, and hard-to-find web sources, so analysts spend their time on sizing, structuring, and credit terms.

This article covers where ChatGPT and Claude help CRE teams today, where they break on underwriting, where borrower files go when they are uploaded, and a checklist to run on any AI tool before trusting its numbers.

Can ChatGPT Underwrite a Commercial Real Estate Deal?

ChatGPT can speed up the work around an underwriting, such as reading long documents and drafting narrative. It should not be the system of record for the numbers. A credit decision needs every row spread, every total exact, and every figure traceable to a page, and general AI tools do not guarantee any of the three.

Definitions Worth Having on Hand

●      Generative AI: AI that produces text, tables, or images from a prompt by predicting likely output. ChatGPT and Claude are general-purpose examples.

●      Deterministic calculation: a calculation that returns the same result every time from the same inputs, using fixed formulas instead of generated text.

●      Reconciliation check: a test that compares figures that should agree, such as line items against a stated total, and flags any difference.

●      Source traceability: the ability to trace any figure back to the specific document and page it came from.

●      Chart of accounts: the standard list of categories used to group income and expense line items. The CREFC chart of accounts is the standard set used in CRE finance reporting.

●      Agentic AI: AI that carries out multi-step work on its own, pulling from documents and external sources and reconciling what it finds, then handing conclusions to a person to decide on.

●      SOC 2 Type II: an independent audit of how a service organization operates its security controls over a period of time.

Where ChatGPT for Real Estate Work Helps CRE Teams Today

Both ChatGPT and Claude are good with language, and a lot of underwriting involves language. These are the places they earn their keep.

Summarizing Long Documents

Appraisals, environmental reports, property condition reports, leases, and loan agreements run long. A general tool can produce a readable first pass in minutes, which helps an analyst decide where to read closely. The summary is a map of the document, and the analyst still reads the sections that drive the credit.

Drafting Narrative Sections

Property descriptions, market overviews, sponsor backgrounds, and the first draft of a risk discussion are writing tasks. General AI drafts them quickly, and an analyst edits for accuracy and tone. Any number inside the draft gets checked against the model.

Answering Quick Questions

What does a cross-default clause do? How do triple-net and gross leases differ? What should I ask a broker about this pro forma? General tools answer questions like these fast, with no setup. Those uses are real, and they are why teams keep using the tools. The trouble starts when a tool moves from reading and writing to producing the numbers a credit decision rests on.

Is Claude Accurate Enough for CRE Financial Analysis?

Claude for real estate work has the same strengths and the same limits as ChatGPT. Both read long files well and both write well. On financial statements, results depend on the file, the prompt, and the day. An underwriting lead at a CRE lender described the experience:

“We've utilized things like Claude and ChatGPT to upload financials and say, transpose this data into our underwriting template and try to help explain what's going on. And more often than not there's been a lot of errors. Those platforms are great, but they're more general purpose applications.”

An underwriting lead at a CRE lender

The experience is common across the industry. Deloitte's 2026 Commercial Real Estate Outlook found that 27% of respondents are experiencing challenges with AI implementation, including technical issues, lack of expertise, or resistance to change, up from 16% a year earlier. General purpose is the key phrase in the quote above. Underwriting asks for things a general tool does not promise, and five gaps matter most.

Skipped Rows and Skipped Properties

A model that reads a financial statement does not count the rows first. If a line is dropped, the output still looks complete, with a subtotal that may or may not match the source. The same pattern shows up in comps. A panelist at the CREFC June 2026 Conference described the first version of an AI agent built to pull them:

“When we did our first iteration, the AI agent skipped any property that didn't have a property name, which as we all know in comps oftentimes the property name is just the address.”

A panelist at the CREFC June 2026 Conference

A skipped record leaves no visible gap, so the only defense is a check built to find it.

Math Errors and Confident but Wrong Numbers

General AI tools generate answers by predicting likely text. That is why they write well, and it is also why a total can read correctly and be wrong. The tool does not pause to say it is unsure. An illustrative example shows how far a single skipped row travels.

Take a 96-unit multifamily T-12 with $1,568,000 in effective gross income and nine expense lines. If one line, real estate taxes at $187,000, is skipped, NOI looks $187,000 higher, and a loan sized to a 9.00% minimum debt yield grows by about $2.1 million.

Illustrative example, not client data.

Nothing in the output flags the problem, and every downstream number inherits it: debt yield, coverage, value, and the loan amount.

No Link Back to the Source Page

A credit committee member asks where the occupancy figure came from. With a general tool, the answer is often a chat transcript. The figure in the model has no link to the page it was read from, so the analyst reopens the PDF and finds it by hand. That re-check is the time the tool was supposed to save.

No Standard Chart of Accounts

Operating statements label the same expense many ways: Repairs and Maintenance, R&M, Building Repairs. General tools map these labels as they see fit, and the mapping can change from one run to the next. Credit teams compare deals, and a portfolio only compares cleanly when every statement lands in the same categories, such as the CREFC chart of accounts.

No Live Market Data

A general model answers from its training data and whatever it can find on the open web. A rent growth assumption or a cap rate comparison needs a governed set of comps with known dates, property types, and submarkets. Search results are not that dataset, and a confident answer does not show how current it is.

What Is the Difference Between General AI and AI Underwriting Software for CRE?

General AI is built to do many things. AI underwriting software is built to produce one thing: numbers a team can defend. AI real estate underwriting works when the tool handles calculations, completeness, sources, standards, data, and security the way a credit committee expects.

Generative AI on Top, Deterministic Calculations Underneath

Smart Capital Center uses generative AI to read documents. The math does not come from generated text. Calculations run deterministically, so the same inputs return the same outputs and an analyst can rerun a spread and get the same answer. Smart Capital Center's agentic AI pulls from borrower files, market comparables, and hard-to-find web sources, and the analyst makes the credit call. Centers Dynamic Partners underwrites deals 4x faster with Smart Capital Center.

Reconciliation Checks That Flag What Does Not Add Up

A reconciliation check compares figures that should agree: line items against a stated total, rent roll income against the T-12, unit counts against the appraisal. Laura Krashakova, Founder & CEO of Smart Capital Center, describes the approach:

“Inside our platform, we recalculate all the numbers and then we raise a flag if something doesn't make sense, if something doesn't add up.”

Laura Krashakova, Founder & CEO, Smart Capital Center

A flag does not decide anything. It tells the analyst where to look.

Every Figure Traced to Its Source Document

Each number in the spread links to the document and page it came from. In credit committee, the answer to “where did this come from?” is a click. The same principle applies to ownership data, covered in CRE Ownership Verification Methods: Traditional vs. AI.

Chart of Accounts, Live Market Data, and SOC 2 Type II

Statements from different borrowers map to the CREFC chart of accounts, so deals compare on the same footing. Benchmarks run against 120M+ properties, 1B+ signals, $500B+ analyzed, and Smart Capital Center is SOC 2 Type II certified. For a wider view of the category, see 20 Best AI Tools for Commercial Real Estate Professionals and how underwriting software platforms compare.

Is It Safe to Upload Borrower Financials to a General AI Tool?

It depends on the account, the settings, and the contract, and many teams cannot answer that question for the tools their analysts already use. A borrower file contains financial statements, rent rolls, tenant names, and ownership details. Once uploaded, it sits on a vendor's systems under that vendor's terms. Jon Carr-Harris, CEO of Cred, put the stakes plainly on a Wharton Scale School panel on AI 2.0 in April 2026:

“If you lose that trust, you lose everything. For anyone in the enterprise space, you can't take data security more seriously.”

Jon Carr-Harris, CEO of Cred, Wharton Scale School panel on AI 2.0, Apr 2026

Where the File Goes After Upload

A file uploaded to a general AI tool leaves the firm's own environment and is processed on the provider's servers. What happens next varies by plan: whether conversations are retained, for how long, whether they can be used to improve models, who at the provider can read them, and where the data is stored. Individual logins and enterprise agreements can carry very different terms.

Questions to Ask Before a Borrower File Leaves Your Environment

●      Is the account an enterprise account under a signed agreement, or an individual login?

●      Is the data retained, and for how long? Can it be deleted on request?

●      Can uploaded files be used to train or improve models?

●      Who at the provider can access the files, and is that access logged?

●      Where is the data stored and processed?

●      Does the provider hold a current SOC 2 Type II report, and will it share it?

●      Does the loan agreement or borrower consent allow the file to go to a third party?

Smart Capital Center is SOC 2 Type II certified, which means an independent auditor has tested how its security controls operate over a period of time.

How Teams Use Both: General AI for Drafting, Purpose-Built Software for the Numbers

A Practical Split

Most teams do not need to pick one. Use general AI where words are the work, and use purpose-built software where numbers go to credit committee.

●      General AI: first-pass summaries of appraisals and reports, narrative drafts, questions to ask a broker or borrower, and plain-language explanations of a clause.

●      Purpose-built software: spreading financial statements, reconciling the rent roll to the T-12, loan sizing inputs and calculations, comps and market benchmarks, and the figures in the credit memo.

One rule keeps the split clean: any number that came from a general tool gets checked against the source before it enters the model. For a longer treatment, see Where Chat Stops and AI in Commercial Real Estate: An Ultimate Guide.

How Should Analysts Check AI Output Before Credit Committee?

Run this checklist on any tool, including purpose-built software. The answers show how much of the file still needs a human check.

1.   Did it read every row? Add the line items yourself and compare the total to the source statement.

2.   Does the math rerun the same way? Run the same file twice and compare the results.

3.   Can I click from each figure to its source page? If not, expect to reopen PDFs in committee.

4.   Which chart of accounts did it use? Confirm each line landed in the right category.

5.   Where did the market data come from, and how current is it? Ask for the comp set with dates, property types, and submarkets.

6.   What happens to the file after upload? Get retention, training, and access terms in writing.

7.   What does it do when something does not add up? A purpose-built tool flags it. A general chat often does not.

8.   Who makes the credit call? The analyst, with the numbers verified.

Fast Drafts, Defensible Numbers: Where the Credit Call Stays With the Analyst

General AI is fast, and underwriting needs numbers a team can defend in credit committee. Smart Capital Center's agentic AI gathers and checks the data, with every number traced to its source, and the analyst makes the credit call. Teams that use ChatGPT and Claude for drafting and Smart Capital Center for the numbers keep the speed and the defensibility.

See every number in your next credit memo trace back to its source document.

Book a demo with Smart Capital Center →

Frequently Asked Questions

Q: Can I use ChatGPT for real estate underwriting?

A: ChatGPT can help with reading documents, drafting narrative sections, and answering quick questions. It should not produce the numbers that go to credit committee, because it does not guarantee complete spreading, exact math, or a link from each figure to its source. Use it for drafts, and use purpose-built software for the numbers.

Q: Is Claude for real estate analysis reliable enough for financial statements?

A: Claude reads long documents well and writes clearly, and results on financial statements vary with the file and the prompt. An underwriting lead at a CRE lender reported that uploading financials to Claude and ChatGPT to transpose into an underwriting template produced errors more often than not. Any figure from a general tool should be checked against the source before it enters a model.

Q: What is AI underwriting software?

A: AI underwriting software reads borrower documents, spreads financial statements, reconciles figures, and builds the analysis an underwriter reviews. CRE-specific versions run calculations deterministically, flag numbers that do not add up, and link every figure to its source. Smart Capital Center's agentic AI works from borrower files, market comparables, and hard-to-find web sources.

Q: How accurate is AI real estate underwriting?

A: Accuracy depends on how the system handles completeness, math, and sources. General AI can skip rows and return confident but wrong totals, and a purpose-built platform adds reconciliation checks and source links to catch those errors. The analyst still reviews the flags and makes the credit call.

Q: Why do AI tools skip rows in financial statements?

A: Generative models predict likely output and do not count the rows in a source before they answer. Long tables, page breaks, merged cells, and scanned pages can make a skipped line more likely, and the output still looks complete. A reconciliation check that adds the line items and compares them to the stated total is what catches it.

Q: Is it safe to upload borrower financials to ChatGPT or Claude?

A: It depends on the plan, the settings, and the contract. Ask whether the data is retained, whether it can be used to improve models, who can access it, and whether the provider holds a current SOC 2 Type II report. Also confirm that the loan agreement or borrower consent allows the file to go to a third party.

Q: How do teams use ChatGPT and Smart Capital Center together?

A: Teams use general AI for summaries, drafts, and quick questions, and Smart Capital Center for spreading, reconciliation, comps, and the figures in the credit memo. Centers Dynamic Partners underwrites deals 4x faster with Smart Capital Center.