How it works
Here Is the Engine Behind It.
A commercial real estate file takes 25 or more hours to underwrite by hand. When somebody says it can be done in under an hour, the correct reaction is suspicion, particularly if you are the person who has to defend the file afterwards. So this page does not ask you to take it on faith. It walks through what actually happens between a loan package arriving and a cited credit memo coming out, what a person still decides, and what we deliberately do not do.
The last line is the one that matters. Nothing here decides the loan.
99.9% extraction accuracy
Dedicated tenant
Never trains shared models
256-bit AES encryption
US-based infrastructure
Start with the honest version
The number sounds inflated until you itemise it. Almost none of that time is credit judgment. It is retyping, reconciling, chasing, formatting and rebuilding. That is why the hour is possible: the work being compressed is not the thinking, it is everything that surrounds the thinking.
Retyping a rent roll into your template, in whatever format the property manager exported it
Tying the rent roll to the operating statement and hunting the differences by hand
Pulling figures out of an appraisal, a T-12, entity financials and a stack of leases
Checking the deal against a policy manual from memory, then arguing about what it says
Assembling the memo, then rebuilding enough of it for the credit officer to form a view
Deciding whether the sponsor is good for it
Deciding whether the market is one you want more exposure to
Deciding whether an exception is worth making, and on what conditions
Deciding how to structure around a tenant concentration you do not love
Sitting in committee and defending the recommendation
Stage one
READINGNot the way a parser reads it. A rent roll is not a table to be scraped, it is tenants, square footage, lease commencement and expiration, base rent, escalations and recovery structure, and every property manager labels those columns differently. The engine knows what it is looking at, which is why it survives a scanned PDF, a handwritten margin note, a multi-tab spreadsheet and a bad photocopy.
Over 70 document types, more than 6,000 extractable data points, and a page-level citation on every value. If a figure cannot be traced to a page, it does not go in the file.
Appraisals, rent rolls, T-12s, profit and loss, personal financial statements, environmental, title, flood.
Stage two
RECONCILINGExtraction on its own is transcription, and transcription inherits whatever the borrower's documents got wrong. The second stage is the one that earns the file: the numbers are cross-checked against each other, and the disagreements are surfaced rather than smoothed over.
This is the step a rushed analyst skips at eleven at night, and it is the step an examiner asks about first.
Stage three
POLICYGeneric underwriting rules are why most lending software gets configured once and argued with forever. LenderBox reads your written credit policy and applies it as written: coverage and loan-to-value floors by property type, concentration limits, guarantor requirements, and the conditions under which an exception is permitted and who has to approve it.
When a deal trips a limit, the file names the provision it tripped and quantifies the gap. It does not approve the exception and it does not bury it.
Dual citation: the provision in your policy, and the page in the borrower's file that trips it.
Stage four
The output is a credit memo in your format with a cited spread underneath it. Your analyst edits it rather than assembles it, and that is the whole point. The judgment, the framing, the recommendation and the signature stay where they were.
In your institution's format, with the exceptions called out and the supporting figures cited to the page they came from.
Every value traceable. A reviewer can check the work instead of taking it on faith, which is a different thing from trusting the output.
Built while the work happens rather than reconstructed afterwards, so the file an examiner reads is the file the analyst worked from.
The part most vendors skip
A list of limits is more useful than another list of features, and if a vendor will not give you one, that is itself the answer.
Approve, decline or price a loan
Overrule your credit policy, or interpret it more generously than it is written
Replace your loan origination system, your core, or your committee process
Invent a number it cannot cite to a page in a document you gave it
Learn from your files in a way that benefits another institution
Remove the assembly work between the documents arriving and the analysis starting
Apply your policy consistently, on the tenth deal of the week the same as the first
Show its work, so a reviewer can disagree with it on the evidence
Make the back book readable, so annual reviews stop starting from nothing
Produce documentation that survives being read by someone outside the institution
Where it runs
Your files sit in your own dedicated tenant. They are not commingled with another institution's, and they are never used to train shared models. That is not a policy position we can change quietly later; it is how the architecture is built, and it is what the SOC 2 Type II attestation was tested against.
Your vendor management team will ask for the evidence long before your lenders see a screen. Read the security overview →
The questions behind the question
Is the hour the whole deal, or just the easy part?
It is the analysis: reading the package, extracting and reconciling the numbers, running your policy and drafting the memo. It is not the relationship, the site visit, the negotiation or the committee. Those were never the 25 hours. What gets compressed is the assembly work, which is where almost all of the manual time actually sits.
How is 99.9% accuracy measured, and what happens to the rest?
Extraction is verified against the source document, and every value carries a page-level citation so the check is available to you rather than asserted at you. That is the point of the citation trail: you are not asked to trust the number, you are given the page. Where a document is genuinely ambiguous, the file surfaces it rather than picking a value and moving on.
What happens when the documents contradict each other?
It gets flagged, not reconciled silently. A rent roll that does not tie to the operating statement, recovery income with no matching lease structure, occupancy in the appraisal that disagrees with the unit-level roll: these are surfaced as disagreements for a person to resolve. Smoothing them over would be the worst possible behaviour for a file that has to be defended.
Is this a black box our examiners will object to?
The opposite is the design goal. Every figure traces to a page in a document you supplied, every policy finding names the provision it came from, and the audit trail is built while the work happens rather than reconstructed afterwards. The file an examiner reads is the file the analyst actually worked from, and a reviewer can disagree with any part of it on the evidence.
Does it learn from our deals, and does that help a competitor?
No. Your documents sit in your own dedicated tenant, are not commingled with another institution's, and are never used to train shared models. Your institution's data makes your own portfolio more useful to you, and stops there.
What does it cost, and how is it priced?
A one-time data activation, a monthly platform fee credited toward usage, and per-deal processing. No annual contract and no per-seat licensing. Setup is not billed, and you begin paying when your team is using the platform on live work. If what we build does not do what we said it would, you get your money back.
What would convince us fastest?
A deal you underwrote last quarter, where you already know the answer. Watch the file get built and check ours against yours. It is a better test than any demo we could script, because the only interesting question is whether it is right on your policy, on your documents.
Reading about an engine is not the same as watching it run. Bring a deal you underwrote last quarter, a review coming due, or the workflow your team keeps rebuilding by hand, and check the output against what you already know.
Set up at no cost. Money-back guarantee. SOC 2 Type II, full report under NDA.
AI-powered commercial real estate lending intelligence. From document intake to committee-ready credit memo in under an hour, end to end.
sales@lenderbox.ai (469) 340-2697
© 2026 LenderBox, LLC. All rights reserved.
