From 15 Hours to 3 Minutes: A CRE Underwriting Workflow, Decomposed | NextAutomation

AI underwriting software does not replace your model or your judgment. What it replaces is the manual chain between a deal landing in your inbox and a scored, memo-ready record: the re-keying, the completeness-checking, the first-pass screen. For one Florida industrial value-add firm, decomposing that chain took the time from intake to a scored record on their mandate from about 15 hours to roughly 3 minutes. That is the honest headline, and the honest caveat comes with it: it is a before-and-after on one firm's specific workflow, not a benchmark you should expect to hit. What ports to your desk is the architecture, not the number.

This guide decomposes that workflow step by step, so you can see what the software does at each stage, where a human still has to decide, and how to measure your own baseline before you believe anyone's time-savings claim. It is written as "here is the workflow you can replicate," because that is the only part of this that transfers.

The Before-State, In Operators' Own Words

Start with the honest baseline, because the result is meaningless without it. At the Florida industrial firm, analysts were spending 15-plus hours per deal re-keying property and rent-roll data out of PDFs and into spreadsheets before any real analysis began. That is not analysis time. It is transcription time, and it scales with deal flow: the busier the firm gets, the more expensive analyst hours disappear into data entry.

If you think that firm is an outlier, listen to a real-estate lending principal describe the identical problem from the debt side: "80 percent of my time is data entry from loan docs into spreadsheets. I read each PDF. I pull out the key data points: principal, interest rate, maturity date, amortization start, borrower covenants, prepayment terms. I type those into a spreadsheet." Different asset, different seat, same bottleneck: a skilled person hand-transcribing structured data that a machine can read.

This is the work AI underwriting software actually removes. Not the thinking. The typing. Everything below decomposes how that removal happens, one step at a time.

Step 1: Intake by Forwarding, Not by Re-Keying

The workflow starts where the deal starts: a broker email with an offering memorandum, a T-12, and a rent roll attached, or a flyer, or a link. In the manual world, receiving that email is the beginning of the 15 hours. In the decomposed workflow, the capture step is forwarding.

You forward the broker's email (or drop the OM into a watched inbox), and the system takes it from there. Nothing is re-typed at this stage. Designing intake this way fits the behavior operators already have: deals arrive as email attachments, so the entry point is an email attachment. A human still reviews what comes out the other side, which we get to in Step 6. But getting a deal into the pipeline goes from a multi-hour transcription job to a forward, and that single design choice is where most of the time in the before-and-after actually comes from.

Step 2: Extraction of the Fields That Matter

Once the document is in, the system parses the attachment and extracts the structured fields: property characteristics, tenancy and rent-roll data, and the financial line items. This is the direct replacement for the lending principal reading each PDF and typing principal, rate, maturity, and covenants into cells. The machine reads the document; the fields land in a structured record.

Be honest about what this step is and is not. Extraction is fast and it is consistent, but it is not infallible. A model can misread a scanned rent roll or miss a field on a messy OM. Anyone who tells you extraction is 100 percent accurate on real-world documents has not run it on enough real-world documents. That is exactly why the next step exists, and why it is not optional. The deep version of this one step, on the two documents that break most often, is covered in our guide to extracting T-12 and rent-roll data with AI.

Step 3: The Completeness Gate

Extraction gives you fields. It does not tell you whether the fields are complete or whether the broker's package was missing something to begin with. That is the job of the completeness gate, and it is the honest backstop that makes the whole workflow trustworthy.

In the Florida firm's setup, every inbound deal runs through a 26-point completeness checklist. The checklist flags what the extraction step could not fill: a missing T-12 month, an absent expense category, a rent roll with no lease-expiration column. This is a completeness gate, not a scoring step, and keeping those two things separate matters. The gate answers "do we have enough to evaluate this deal," not "is this a good deal." When a field is missing, it surfaces as a flag a person can act on, rather than silently passing a half-empty record downstream to be scored on data that is not there.

This is also the honest answer to the operator who asked us, in plain words, "when documents are missing from a client, where does it show for me?" It shows here, as an explicit flag on the record, before the deal is scored or a memo is drafted. The gate is what lets you trust the speed.

Step 4: Scoring Against Your Mandate

Only once a record is complete does it get screened. In the decomposed workflow the screen runs on 5 weighted criteria plus 3 automatic disqualifiers, and the critical detail is that these are configured per mandate. They are not a universal grade. A value-add industrial buyer and a stabilized multifamily buyer do not share a buy box, so they do not share a screen.

That is the whole reason this step is configurable rather than hard-coded. The automatic disqualifiers are the fast "no" (wrong market, wrong asset class, a deal-breaker on basis) that lets a person stop reading immediately. The weighted criteria produce a first-pass ranking so the deals worth a human's attention rise to the top of the inbox; we lay out how to design that screen in our CRE deal screening playbook. What the system does at this step is screen, score, and rank. What it explicitly does not do is make the investment decision. It is triage, not a verdict, and treating the score as a verdict is the fastest way to misuse a tool like this.

Step 5: Memo Generation From a Single Snapshot

A scored record is useful internally, but the artifact a committee actually reads is a memo. The last automated step turns the structured snapshot into documents: one upload in, and out comes a deal snapshot, an investment-committee memo draft, and an asset business plan. On the memo side, the same firm went from days of analyst work assembling an IC package to a branded PDF and an editable deck generated on demand.

Two honesty points hold this step together. First, the analyst still supplies the model. The document generator does not underwrite the deal; it assembles the narrative and the exhibits around the numbers the analyst produces. Second, because the memo and the business plan generate from the identical snapshot, they never disagree with each other, which is a documentation-integrity property, not an accuracy claim about the deal itself. If you want the full decomposition of just this stage, we cover it in going from OM to IC memo with AI.

Step 6: The Human Review That Is Not Optional

This is the step vendors quietly skip in their demos, and it is the step that makes the workflow legitimate. In the Florida firm's process, the investment committee never reads raw AI output. It reads a document an analyst has reviewed and approved. The human review is a built-in stage of the workflow, not a nice-to-have bolted on afterward.

Put the whole chain together and the division of labor is clear: the software captures, extracts, gates for completeness, scores against the mandate, and drafts the memo. A person reviews the completeness flags, sanity-checks the extracted fields, supplies the model, and approves the document before it goes to committee. The 15-hours-to-3-minutes compression happens in the mechanical middle of that chain. The judgment at the ends stays human, on purpose. That is not a limitation to apologize for; it is the reason the output is safe to underwrite on.

What You Can Actually Replicate (and What You Cannot)

Here is the honest line between the two. The architecture replicates. The number does not.

The six-step structure, forward-in intake, extraction, a completeness gate, mandate-specific scoring, single-snapshot memo generation, and a mandatory human review, ports across asset classes. The same skeleton works for multifamily, retail, and office; the extraction targets and the scoring criteria are what you swap. What does not port is the 15-hours-to-3-minutes result, which is a before-and-after on one Florida industrial value-add firm's specific mandate and document set. Sometimes you will see the same measurement quoted as "roughly 300 times faster," which is simply that identical intake-to-scored-record number expressed as a ratio, not a second, independent proof point and not a claim about your deals.

Your number will be different, because it depends on your before-state, your document quality, and how tightly your mandate can be encoded. That is the point of decomposing the workflow rather than selling the headline: you adopt the structure, then measure what it actually does for you. The broader adoption context is real, for what it is worth. JLL's 2025 Global Real Estate Technology Survey (published October 28, 2025, of 1,500-plus decision-makers across 16 markets) found 88 percent of investors and owners now piloting AI, up from under 5 percent in July 2023. But an industry piloting AI is not a validated benchmark for your desk, and you should treat it that way.

How to Measure Your Own Baseline

Before you can know what any AI underwriting software saves you, you have to know your before-state honestly. Most firms never write it down, which is exactly why vendor time-savings claims are impossible to sanity-check. Here is how to build your own baseline.

• Time the intake-to-scored-record leg specifically. Do not measure "underwriting" as one blob. Measure the exact leg this workflow compresses, from the moment a broker's deal lands to the moment you have a complete, screened record. That is the number that moves, and the one the Florida firm's before-and-after describes.
• Count the re-keying separately from the thinking. How many analyst hours per deal are pure transcription of PDF data into your model, versus actual judgment? The lending principal's "80 percent" was all transcription. Splitting these tells you your realistic ceiling, because the software removes the first and leaves the second.
• Write down your mandate as rules. If you cannot state your buy box as a short list of weighted criteria and hard disqualifiers, no scoring step can encode it. The clarity of your mandate is the ceiling on how much of the screen can be automated.
• Log your completeness failures for a month. How often does a broker package arrive missing a field you need? That rate tells you how much work the completeness gate will actually save, and it is usually higher than firms guess.

Do this for a handful of live deals and you have a defensible before-state, so any after, ours or anyone else's, becomes measurable instead of a slogan. To compare tools against this decomposition, we keep a running evaluation in our guide to the best AI tools for CRE underwriting, and the wider market picture is in the state of AI in commercial real estate in 2026.

Frequently Asked Questions

Is the 15-hours-to-3-minutes number a benchmark I should expect to hit?

No. It is a before-and-after on one Florida industrial value-add firm's specific mandate and document set, measuring the intake-to-scored-record leg of their workflow. It is not a cross-firm benchmark and not a promise. What replicates is the six-step architecture behind it; your own result depends on your before-state, your document quality, and how cleanly your mandate can be encoded. The point of decomposing the workflow is so you can adopt the structure and then measure what it actually does for you, rather than trusting a headline.

What does AI underwriting software actually replace?

The mechanical chain between a deal landing in your inbox and a scored, memo-ready record: the re-keying of PDF data into spreadsheets, the completeness-checking, the first-pass screen, and the assembly of the memo draft. It does not replace the model or the investment decision. An analyst still supplies the underwriting model, reviews the extracted fields and completeness flags, and approves every document before it reaches committee. The software removes the typing, not the thinking.

Does the completeness gate score the deal?

No, and keeping them separate is the point. The completeness gate (a 26-point checklist on every inbound deal in the workflow described here) answers "do we have enough data to evaluate this deal" by flagging missing fields, such as an absent T-12 month or expense category. Scoring is a separate, later step that answers "does this deal fit our mandate," using weighted criteria and automatic disqualifiers configured per mandate. Running the gate first is what stops a half-empty record from being scored on data that is not actually there.

Does this workflow work for asset classes other than industrial?

The architecture ports across multifamily, retail, and office. The extraction targets and the mandate-specific scoring criteria are what you swap for each asset class; the six-step skeleton (intake, extraction, completeness gate, scoring, memo generation, human review) stays the same. What does not port is the specific 15-hours-to-3-minutes number, which belongs to the one industrial firm it was measured on. Expect the same structure to help across asset classes and expect your own time figure to be different.

Where do I see it when a broker's package is missing a document?

On the deal record itself, as an explicit flag, before the deal is scored or a memo is drafted. That is the job of the completeness gate in Step 3: it checks every inbound deal against a fixed checklist and surfaces exactly what is missing, so a person can chase the missing rent roll or T-12 month rather than discovering the gap after underwriting on incomplete data. The flag is the mechanism that lets you trust the speed of the earlier steps.