Underwrite commercial real estate faster with AI

To underwrite commercial real estate faster with AI, give it the document work and keep the judgment. Extraction, tie-outs between the seller's documents, and data entry into the template you already use can go to the model, while rent growth, exit cap, and downtime remain the analyst's call.

That is a narrow claim on purpose, and we think it is the useful one for a capital markets or investment sales team that already has a model it trusts.

Where do the hours go in a first-pass underwrite?

Your model already does the math. Once the inputs are in the right cells, the cash flow, the debt, and the returns calculate themselves.

The first day of a deal goes to something else. An analyst opens the offering memorandum, a rent roll exported from the seller's property management system, and a trailing twelve month operating statement, and none of them share a format. The rent roll might be a PDF with merged cells. The T-12 uses the seller's account names, which do not match your chart of accounts. The analyst keys all of it in, and along the way notices the things that do not line up. That noticing is the valuable part of the day, and it tends to get squeezed by the typing.

What AI is good at here, and what stays with the analyst

Work	Who does it
Reading the OM, rent roll, and T-12 and pulling values into a standard table	AI, with a source page on every row
Mapping the seller's T-12 lines to your chart of accounts	AI drafts the mapping, the analyst approves it once per deal
Tie-outs between documents	Written checks that run the same way every time
Writing inputs into the template	AI, restricted to input cells
Rent growth, exit cap, downtime, leasing costs, hold period	Analyst
Deciding whether a discrepancy matters	Analyst
Sign-off on what goes to a client or committee	Analyst and broker

Step one: pull the rent roll and T-12 into a fixed table with a source page for every row

The model should never write straight from a PDF into your underwriting model. It should fill an intermediate table first, with the same columns on every deal: unit or suite, tenant, square feet, rent, lease start, expiry, escalations, and reimbursements. A second table holds the T-12, one row per line item, with the seller's label, the amount, and the account in your chart of accounts it maps to.

Every row carries the page it came from. Claude's API documentation lists converting document information into structured formats as a use case for its PDF support and says it reads text, pictures, charts, and tables. The same page gives limits worth knowing before you send a full due diligence folder: a maximum request size of 32 MB, which it notes varies by platform, and 600 pages per request, or 100 pages when the request's context window is under 1 million tokens.

Two instructions matter more than the rest. Tell the model to use only the documents provided, and tell it to leave a cell blank and say so when a value is not stated. Anthropic's guidance on reducing hallucinations recommends both, along with grounding answers in direct quotes from the source. A blank cell with a note is quick for the analyst to resolve, and a plausible invented escalation is much harder to catch later. The same discipline applies when you abstract a commercial lease with AI, and the abstract is often what settles a question the rent roll leaves open.

Step two: tie-out checks before anything touches the model

An experienced analyst runs a set of checks in their head while keying a deal. Those checks should be written down as arithmetic that runs the same way every time, in a script or in formulas on the intermediate table, and they should not be left to the model's judgment. A typical set:

• Annualized rent from the rent roll against rental income on the T-12, with the difference shown.
• Total square footage on the rent roll against the figure in the OM.
• Unit or suite count against the source document's own count.
• Occupancy implied by the rent roll against the occupancy the OM states.
• Every lease expiring inside your hold period, listed with tenant and square feet, whether or not the OM mentions it.
• T-12 lines that did not map to an account, and whether the mapped lines sum to the seller's stated net operating income.

A failed check does not mean the extraction was wrong. Sellers' documents disagree with each other for ordinary reasons, such as a rent roll dated after the T-12 period ends. The check puts the gap in front of the analyst on the first morning of the deal, which is when it is cheapest to ask the listing broker about it.

Step three: populate your own Excel template with formulas intact

Only after the tie-outs are reviewed do values move into the model. The write is limited to the input cells or named ranges your template already defines, on a copy of the master file, so every formula stays the one your team wrote.

Anthropic's documentation for Claude for Excel says the add-in can populate existing templates, updates cell values "while keeping formula relationships intact," and warns before overwriting existing data. It also lists data tables, macros, and VBA operations as unsupported, so a template that depends on a macro to roll the rent schedule, or on a data table for its sensitivity grid, needs those steps run by a person or rebuilt.

Because each input traces back to a row in the intermediate table, and each row to a page, the analyst can answer "where did this number come from" for any cell.

Putting the seller's assumptions beside the in-place numbers

The OM comes with its own pro forma. A useful output of the same workflow is a short comparison table: in-place rent beside the seller's market rent, T-12 expenses beside the seller's year-one expenses, in-place occupancy beside stabilized occupancy. The model can lay those out. It should not choose between them. Seeing the gap in one place makes the analyst's own assumptions faster to set and easier to defend.

Working inside Excel versus uploading files to a chat window

Both have a place. Extraction from PDFs is a document-reading task and does not need the workbook open. The work in the model itself is better done inside the workbook. According to the Claude for Excel documentation, the add-in answers questions about the open workbook with cell-level citations you can click, which suits the review step: an analyst can ask how a number was derived and land on the cell.

If your team repeats this on every deal, the instructions, column definitions, and account mapping belong in a packaged skill so each analyst gets the same behavior. We cover that in Claude skills for commercial real estate.

What does the analyst sign off on?

Anthropic states that Claude for Excel is not recommended for "final client deliverables without human review" or "audit-critical calculations without verification." We agree with the vendor, and we would design the review in even if the documentation were silent.

A workable sign-off covers four things: the tie-out report, with each variance explained or sent back to the seller; the account mapping; a sample of extracted rows checked against their source pages, weighted toward the largest tenants and the nearest expirations; and every assumption, which the analyst sets personally. For more on structuring that step, see how to design the human review step in an AI workflow.

Seller-supplied files are untrusted input

The same documentation carries a warning that applies directly to deal work: "Only use Claude for Excel with trusted spreadsheets. Files from external sources can contain hidden instructions that manipulate the add-in into extracting data, modifying records, or performing destructive actions."

A seller's Excel rent roll is an external file. In practice that means the workflow reads seller files as data and copies values into your own clean table, the add-in works in your template and not in the seller's workbook, and anyone running it reads the confirmation prompts before approving a change. The documentation also says the add-in is not recommended for models containing highly sensitive or regulated data without proper controls, and seller files usually arrive under a confidentiality agreement. Your IT or security lead and your counsel should confirm how this fits your firm's policies and the terms you signed.

Dedicated underwriting software or a workflow built on your model: how to decide

Dedicated platforms have real strengths. An Altus Group article from August 2026 reports the view that general-purpose models and specialized software are complementary, that an AI tool "has to be right the same way twice," and that "the accuracy of AI outputs is only as good as the data underlying them." Altus sells valuation software, so read that as an informed industry position. The reproducibility point is the reason step two is written as fixed checks.

The questions that decide it are practical. Consider whether your team would adopt a vendor's model or keep its own, whether the investors and lenders you send models to expect your format, and whether you can trace each input to a source page in either setup. If the answers favor your own model, the three steps above are the build. Any time saving a vendor quotes is worth testing on three of your own recent deals before you believe it, and the same goes for this method.

Most firms are still early on this. In Deloitte's 2027 commercial real estate outlook, a survey of 950 executives and their direct reports, 92% of respondents were in the piloting or research phase with AI and 8% had integrated AI solutions. Document extraction into Excel and BOV first pass work are among the things we build with brokerage teams, and if you want to walk through your own template, you can get in touch.

Common questions

Can AI underwrite a commercial real estate deal?

It can do the document work of a first pass: extract the rent roll and T-12, flag where the seller's documents disagree, and populate your template for review. The assumptions and the conclusion remain the analyst's, and Anthropic's own documentation says Claude for Excel is not recommended for final client deliverables without human review.

How do I extract a rent roll from a PDF into Excel with AI?

Give the model the PDF and a fixed list of columns to fill, one row per unit or suite, and require the source page number for each row. Have it leave a cell blank when the document does not state a value, then check the row count and totals against the PDF before the table goes anywhere near your model.

Will AI break the formulas in my underwriting model?

Anthropic's documentation says Claude for Excel updates cell values while keeping formula relationships intact and warns before overwriting existing data, and it also lists data tables, macros, and VBA operations as unsupported. The safe design is to limit writes to named input cells on a copy of the template and compare the formula cells to the master afterward.

How do I check AI-extracted numbers before relying on them?

Use tie-outs that do not depend on the model's judgment: the rent roll total against T-12 rental income, square footage against the offering memorandum, and unit or suite count against the source. Then spot check individual rows against the page number recorded beside each one.

Is AI underwriting accurate enough for an investment committee memo?

Treat the output as a first pass that an analyst has to review line by line. Anthropic's guidance on reducing hallucinations says its techniques do not eliminate errors entirely and that critical information should always be validated for high-stakes decisions.