An AI-assisted underwriting workflow is not a machine that decides deals for you. It is a pipeline that takes a broker’s offering memorandum out of your inbox, turns its numbers into a structured model, flags the figures that do not tie out, and hands your analyst a first-pass return to interrogate — so the person who owns the capital decision spends their time on judgment instead of data entry. The difference between a firm that gets value from this and a firm that gets a mess is entirely about where the line between machine and human is drawn, and at a 4–20 person shop that line has to be drawn deliberately. This is the stage-by-stage anatomy of what actually happens between the email and the decision, what the AI does at each step, and the one step it must never own.
The five stages, end to end
Strip away the vendor language and every AI-assisted underwriting workflow, whether it runs on a purpose-built platform or a spreadsheet and a chat assistant, is the same five stages in the same order. Naming them is what lets you see where the automation helps and where a person still has to stand.
| Stage | What happens | Who owns it |
|---|---|---|
| 1. Intake | A deal arrives, gets identified, and is triaged in or out | AI drafts, human confirms |
| 2. Extraction | OM, rent roll, and T-12 become structured numbers | AI does the work |
| 3. Reconciliation | The numbers are checked against each other for what does not tie out | AI flags, human reads flags |
| 4. First-pass model | A return is calculated on stated assumptions | AI drafts |
| 5. Decision | Assumptions are challenged and the deal is advanced or killed | Human owns it |
The mistake firms make is treating stage 4 as the finish line — as if a computed return were an answer. It is a draft. The value of the whole pipeline is that it gets a defensible draft in front of a person fast, so the person can spend their scarce hours on the part no model should do: deciding whether the story the deal tells is true.
Stage 1: Intake — from inbox to a triaged deal
The workflow starts where the deal does: an unsolicited email from a broker with an offering memorandum attached, one of dozens that land in a given week. Before any underwriting happens, the deal has to be identified — property, asset class, market, asking price, broker — and triaged against your buy box. Most of what hits the inbox does not fit, and the point of intake is to spend zero analyst minutes on the deals that were never going to clear.
Here the AI reads the email and the attachment and produces a short structured summary: what the asset is, where it is, what is being asked, and how it maps to the criteria you have told it to watch. This is exactly the job Dealpath’s AI deal-screening feature is built to do at the platform level — read the incoming documents and produce a deal summary and next-step suggestion in seconds. A lean firm can get the same first cut from a general assistant reading the PDF. Either way, the machine is drafting a triage note; a person is glancing at it and deciding whether the deal earns a real look. Screening the many so you underwrite the few is the whole discipline, and we make the fuller case for it in the deal-screening framework for a lean team.
The failure mode this stage catches is the one every small shop knows: the good deal buried under two hundred broker blasts, missed because no one had time to open the attachment.
Stage 2: Extraction — documents into structured numbers
Once a deal is worth underwriting, the grind begins — and this is the grind AI actually removes. An offering memorandum, a rent roll, and a trailing twelve-month operating statement arrive as inconsistent PDFs, sometimes scanned, sometimes exported cleanly, never in the same format twice. Underwriting cannot start until those documents become structured numbers sitting in the cells of your model: unit-by-unit rents, lease expirations, reimbursements, line-item operating expenses, in-place net operating income.
This has always been the least valuable, most time-consuming part of an analyst’s day — retyping a rent roll into Excel, keying a T-12 line by line. It is also the part a current-generation model does well. The AI reads each document, pulls the figures, and maps them into the model’s structure. The best implementations do one thing that matters more than speed: they cite every extracted number back to the page and line it came from, so the figure in your model is traceable to the source document rather than trusted blindly.
The extraction is not the point of the workflow, and treating it as the point is the trap the vendor pitches set. Getting clean numbers into the model quickly is table stakes. What you do with those numbers next is where the deal is actually screened.
Stage 3: Reconciliation — the step that earns its keep
Here is the stage the product demos skip, and it is the one that justifies the whole pipeline. You now have three views of the same asset — the rent roll, the trailing operating statement, and the broker’s pro forma in the OM — and they do not agree. They never do. The rent roll shows in-place rents; the OM projects a stabilized number that assumes every unit turns to a higher mark. The T-12 shows real operating expenses; the pro forma quietly trims a few line items to make the yield look better. Broker optimism lives in the gap between these documents.
Reconciliation is the machine reading all three and flagging what does not tie out: the pro forma income that sits well above trailing actuals, the vacancy assumption below anything the market supports, the expense line that vanished between the T-12 and the pro forma. It is not deciding whether those gaps are justified — some are, a genuine value-add plan explains a lot — but surfacing them so a person looks in the right place. This is the difference between a tool that speeds up data entry and a tool that changes how you underwrite. The extraction saves an hour; the reconciliation catches the deal that was dressed up to look like something it is not.
For a firm coming from a spreadsheet, this is the capability worth reaching for first, and it is achievable well short of a six-figure platform. We walk through where a general assistant handles this and where it stops in Excel plus ChatGPT versus a custom underwriting copilot.
Stage 4: First-pass model — returns, not a verdict
With reconciled numbers in place, the workflow computes a first-pass return: an unlevered yield, a levered return on stated financing assumptions, a going-in cap rate against the ask. On a platform this is native; with a general assistant it is your existing model doing the math on the numbers the AI populated. Either way the output is a set of returns tied to an explicit list of assumptions — rent growth, exit cap, hold period, financing terms.
The word to hold onto is stated. The model has calculated what the return would be if the assumptions were true. It has not decided whether they are. A first-pass return that clears your threshold is not a green light; it is a prompt to ask why. A return that fails is not automatically a kill; it might just mean the broker’s exit cap is aggressive and yours is conservative. The model’s job is to get a defensible number and its full assumption set in front of a person quickly — which is precisely the setup for the only stage a human can own.
This is also the boundary where the honest limits of automation sit. A tool whose numbers feed a screening flag can tolerate the occasional miss because a person confirms every advance. A tool whose numbers flow straight into an investment-committee memo needs far more validation, because a quietly miscalculated cap rate is worse than no tool at all. Which of those you are building reshapes everything downstream, and it is a decision to make before you commission anything.
Stage 5: The analyst owns the call
The last stage is the one the machine never touches: someone looks at the reconciled model, the flagged gaps, and the first-pass return, and decides whether the deal tells a true story. They challenge the assumptions the AI accepted — is that rent growth real for this submarket, does the exit cap hold, is the value-add plan financeable — and they advance the deal to a full underwrite or kill it.
This is not a courtesy checkpoint you could automate away next year. It is the design. Underwriting decisions carry real capital risk, and a model has no accountability for a bad one. The correct architecture for a lean firm keeps the AI on the volume — intake, extraction, reconciliation, the arithmetic of the first pass — and keeps the person on the judgment. Traditional institutional modeling tools like Argus have always assumed a skilled human driving them; AI does not change that, it just removes the keystrokes between the documents and the decision. We cover where each approach fits in Argus versus AI-assisted underwriting.
Drawn this way, the workflow does not replace your analyst. It gives one analyst the reach of three, because the two hours that used to go into keying a rent roll now go into pressure-testing the deals worth pressure-testing.
Where to run this: assistant or platform
The same five stages run on very different setups, and a small firm should not assume the answer is a purpose-built platform. A firm that underwrites a handful of deals a week can run most of this workflow today with a business-tier ChatGPT or Claude subscription reading its documents into an existing Excel model — no new system, no integration project. The assistant handles intake summaries, extraction, and a first cut at reconciliation; the analyst does the rest in the spreadsheet they already trust.
The purpose-built platforms — Dealpath, and lender-focused tools like Blooma — earn their cost when volume climbs, when several people need the same pipeline in front of them, or when the deal data has to write back into a shared system rather than living in one analyst’s workbook. The honest on-ramp is to prove the workflow with the tools you have, then graduate to a platform when the bottleneck is genuinely coordination and volume rather than keystrokes. What that build actually costs, line by line, is covered in our guide to what custom underwriting automation costs, and where this whole capability sits in a lean firm’s operating rhythm is the subject of the CRE deal-analysis playbook.
Handling confidential deal data
One stage-zero question decides which tools you may use at all: an offering memorandum and a rent roll are confidential, sometimes under a non-disclosure agreement, and you are about to paste them into an AI system. For a firm with no IT department, the rule is to use business or enterprise tiers of the major assistants, which by policy do not train on your inputs, rather than free consumer tiers. Purpose-built CRE platforms handle this at the account level, but the same diligence applies — read what the vendor does with your documents before the first OM goes in. This is not a reason to avoid the workflow; it is a reason to choose the tier deliberately. The broader argument for why a small firm can run this kind of owned capability and out-operate larger competitors by doing so is the case we make in the small CRE firm AI manifesto.
FAQ
What is an AI-assisted underwriting workflow?
It is a five-stage pipeline that runs a deal from inbox to decision: intake triages the deal against your buy box, extraction turns the offering memorandum, rent roll, and T-12 into structured numbers, reconciliation flags where those documents disagree, the model computes a first-pass return on stated assumptions, and a person challenges the assumptions and decides. AI does the first four stages; a human owns the fifth. The point is not to automate the decision but to get a defensible draft in front of an analyst fast.
Does AI actually decide which deals to buy?
No, and any workflow designed that way is built wrong. AI handles the volume work — reading documents, populating the model, flagging inconsistencies, calculating returns on the assumptions it was given. Whether those assumptions are true, and whether the deal is worth pursuing, is a judgment that carries capital risk and stays with a person. The correct design keeps the machine on keystrokes and the human on the call.
Which stage delivers the most value?
Reconciliation, not extraction. Extraction saves time by removing manual data entry, which is real but commoditized. Reconciliation — the machine reading the rent roll, trailing statement, and broker pro forma together and flagging what does not tie out — is where the workflow changes outcomes, because that gap is exactly where broker optimism and errors hide. A tool that only extracts is a faster typist; a tool that reconciles changes how you screen.
Can a small firm run this without buying a platform?
Yes. A firm underwriting a handful of deals a week can run most of the workflow with a business-tier ChatGPT or Claude subscription reading documents into an existing Excel model. Intake summaries, extraction, and a first pass at reconciliation are all achievable without a new system. Purpose-built platforms like Dealpath earn their cost at higher volume, when multiple people need a shared pipeline, or when data must write back into a common system.
What documents does the workflow read?
The core three are the offering memorandum, the rent roll, and the trailing twelve-month operating statement. The OM carries the broker’s pro forma and narrative; the rent roll carries in-place rents and lease terms; the T-12 carries real operating history. Some workflows also ingest appraisals, budgets, or broker opinions of value. Reading all three together is what makes reconciliation possible.
How does AI avoid making up numbers during extraction?
The safeguard is source citation: strong implementations tie every extracted figure back to the page and line it came from, so a number in the model is traceable to the source document rather than trusted blindly. That lineage lets an analyst spot-check anything that looks off and is the difference between an auditable model and a black box. It does not remove the analyst’s responsibility to verify, but it makes verification fast.
Is it safe to put confidential deal documents into an AI tool?
It can be, with the right tier. Use business or enterprise tiers of the major assistants, which by policy do not train on your inputs, rather than free consumer tiers, and read any purpose-built platform’s data-handling terms before uploading an OM. Because offering memoranda are often under an NDA, the tool choice is a compliance decision, not just a feature one — but it is a solvable one, not a reason to avoid the workflow.
How much time does this actually save?
Most of the saving comes from removing the manual data entry that has always dominated an analyst’s underwriting hours — keying a rent roll and a T-12 into a model before any thinking begins. The workflow does not make the judgment faster; it removes the hours in front of it, which lets one analyst give a real look to far more deals in a week. The gain is reach and consistency, not a shortcut on the decision itself.
When should a firm graduate from an assistant to a purpose-built platform?
When the bottleneck stops being keystrokes and starts being coordination: several people needing the same live pipeline, deal data that has to write back into a shared system, or a deal volume high enough that a spreadsheet-and-assistant setup can no longer keep everyone current. Below that, a platform adds cost and integration work without solving a problem you have yet.
Does this replace an analyst or a modeling tool like Argus?
Neither. It removes the keystrokes between the documents and the model, which is what frees an analyst to underwrite more deals well. Institutional modeling tools like Argus still assume a skilled person driving them; AI shortens the path to a populated model but does not take over the modeling judgment. The workflow augments the analyst and feeds the model — it replaces the data entry, not the professional.
Key takeaways
- An AI-assisted underwriting workflow is five stages — intake, extraction, reconciliation, first-pass model, decision — and AI owns the first four while a person owns the fifth.
- Reconciliation, not extraction, is where the value is: the machine flagging where the rent roll, T-12, and broker pro forma disagree is what catches the deal dressed up to look better than it is.
- A first-pass return is a draft on stated assumptions, not a verdict; the human stage exists to challenge those assumptions, and no responsible design automates it away.
- A small firm can run most of the workflow on a business-tier assistant and its existing Excel model, graduating to a purpose-built platform only when volume and coordination — not keystrokes — become the bottleneck.
- Confidential documents make tool-tier choice a compliance decision: use business or enterprise tiers that do not train on your inputs, and read a platform’s data terms before the first OM goes in.
Want to know which stages your firm should automate first, and whether an assistant or a build fits your deal flow? A short conversation about your documents, your volume, and the systems you already run will map it far better than any template. Book your free AI-readiness assessment → and we will size where an AI-assisted underwriting workflow would earn its place in your process.
Dirk Jan van Veen, PhD
