Let AI do the reading, the reconciling, and the first draft of the memo — never the numbers you have not checked and never the assumptions that set the price. AI-assisted underwriting works when a person who owns the capital decision spends their hours on judgment instead of transcription, and it fails the moment a clean-looking figure the model pulled off the wrong column of a T-12 quietly moves a bid. The gap between those two outcomes is not the tool. It is the discipline you bring to it. These ten rules are the code an underwriter at a lean firm can follow with any tool, or with none, to capture the speed without inheriting the failure mode. They end in a one-page rulebook you can pin next to the model.
What AI-assisted underwriting actually is
AI-assisted underwriting uses a model to do the mechanical work of turning a deal package into an analysis — reading the offering memorandum, pulling the rent roll and the trailing twelve into structured fields, checking the numbers against each other, and drafting a first-pass return — while a person owns the parts that determine whether the deal is real. It is assistance, not replacement. The word “assisted” is the whole point, and firms that drop it get burned.
The distinction that matters is between the work that is tedious and the work that is judgment. Typing a scanned rent roll into a model is tedious; deciding whether the exit cap is defensible is judgment. AI is very good at the first and dangerous at the second, because it produces both with the same fluent confidence. The rules below draw that line and keep it drawn under deal pressure. They assume you already screen your inbox before you underwrite — if you do not, start with the discipline of screening every deal and underwriting only the survivors, because underwriting a deal that never had a chance is the most expensive habit a lean team keeps.
Rule 1: Screen before you underwrite
Underwriting is expensive attention. Spend it only on deals that have already cleared a cheap first cut against your buy box. A firm that puts every broker blast through a full model gives the two deals worth winning the same tired half-day as the thirty that were dead on arrival — and the good one gets underwritten badly because the analyst is spent by the time it arrives.
AI belongs at both stages, but they are different jobs. Screening ranks the inbox so you know what to open; underwriting models the cash flows so you can price a bid. Use AI to triage wide and cheap, then bring the full discipline of these rules to the few deals that survive. The complete two-stage funnel is laid out in our deal-analysis playbook.
Rule 2: Let AI extract, never let it decide the number is right
Extraction is where AI saves the most time and hides the most risk. A model will pull the in-place rents, the expense lines, and the NOI out of a messy PDF in seconds — and it will do it correctly almost every time, which is exactly what makes the rare miss dangerous. The errors that hurt are not the obvious ones. They are the figure read off the adjacent column of a financial table, or the right formula applied to a variable the model extracted wrong: a lease commencement date off by a digit, an expense that quietly went missing, a “total” that is actually a subtotal.
These pass review because they look finished — a well-structured number in a clean table does not announce that it came from the wrong row. Treat every extracted figure as a claim to verify, not a fact to accept. The model’s job is to propose the number and show you where it found it; your job is to confirm the ones a decision hinges on. Never let extraction speed talk you out of the check.
Rule 3: Make every figure cite its source
The rule that makes Rule 2 practical: require grounding. Every number the model puts in the analysis should link back to the page and line it came from, so a person can verify in one click. A figure with no traceable source is a rumor, not data, and a model that cannot show its work is asking you to trust it on exactly the thing it is worst at.
This is the single most valuable habit in AI-assisted underwriting. When the trailing twelve says NOI is one number and the OM proforma says another, grounding lets you see instantly which document each figure came from and which one the broker is asking you to believe. Whether you run a purpose-built platform or a general assistant reading the PDF, insist on source-linked figures; a tool that cannot provide them is not ready to touch your numbers.
Rule 4: Reconcile the documents against each other
The value of AI-assisted underwriting is not the raw extraction — it is the cross-check. A deal package is a set of documents that should agree and frequently do not: the OM’s stated occupancy against the rent roll, the proforma NOI against the trailing twelve, the tax line against the county record. Every gap is either an honest artifact of timing or a story the seller is telling, and finding the gaps is where a lean firm gets the most out of the machine.
Point AI at the reconciliation explicitly: ask it to list every figure that does not tie out across the documents and to name which source each version came from. The model is fast and tireless at this comparison and will surface discrepancies a rushed analyst misses at eleven at night. It does not resolve them — you do — but a list of what does not agree is the most useful thing AI produces in the whole workflow. For the full stage-by-stage view, see our anatomy of an AI-assisted underwriting workflow.
Rule 5: Own every assumption yourself
Extraction is about the past — what the property did. Assumptions are about the future — what you believe it will do — and they set the price. Exit cap, rent growth, downtime, renewal probability, capital reserves, the pace of a lease-up: these are the levers that decide the bid, and they are pure judgment about a specific market at a specific moment. This is the line AI must not cross.
A model will happily suggest an exit cap and a rent-growth curve, phrased with the same confidence it uses for a number read straight off the page. That confidence is the trap: it has no view of the submarket’s next eighteen months, no read on the seller’s motivation, no memory of the deal that traded down the street last quarter. Set your own assumptions from your market knowledge and defensible comparables, and let the model calculate the return on them. The discipline behind those comparables — where AI helps and where it cannot — is the subject of our comp selection framework for an AI-assisted world. If you take one rule from this list, take this one: AI models the math; you own the assumptions.
Rule 6: Keep the go/no-go human by design
The model reads, reconciles, and drafts. A person decides, and capital never moves on an analysis no human has interrogated. Build the workflow that way on purpose, because the failure mode of a good AI underwriter is not that it breaks — it is that it works well enough that people stop checking, and the stop is invisible until the deal that needed a second look did not get one.
Draw the line where the analyst reads the reconciled model, challenges the assumptions, verifies the figures that matter, and makes the call. That is not distrust of the tool; it is the correct division of labor, and it is the same principle that lets a lean shop out-operate a bloated one — AI clears the busywork so human judgment lands where it counts, as the small-firm manifesto argues. A vendor who tells you the tool replaces the underwriter is describing a liability, not a feature.
Rule 7: Underwrite from a template, not a blank prompt
Consistency is what makes AI output trustworthy over a pipeline of deals. A blank prompt produces a different analysis every time — different fields, different order, different assumptions surfaced or buried — and you cannot compare deals underwritten inconsistently. A standardized template fixes the shape: the same extracted fields, the same reconciliation checks, the same assumption slots left explicitly empty for you to fill, the same first-pass return format.
Write your underwriting standard once — what a first pass must contain and in what order — and make every deal run through it, whether the engine is a platform or an assistant reading a structured prompt. The template does two things: it makes the output predictable enough to review quickly, and it forces the assumptions into the open so no exit cap slips in unexamined. A template is prompt discipline made permanent.
Rule 8: Keep an audit trail of what the model touched
When AI touches a model, you need to know later which numbers it pulled, which it flagged, and which a person changed. An investment committee, a lender, or your own future self will ask where a figure came from, and “the AI produced it” is not an answer you can defend.
Keep it simple: preserve the source-linked extractions, note where you overrode a figure or an assumption, and keep the reconciliation flags with their resolutions. This is ordinary underwriting hygiene, but AI raises the stakes because it produces so much so fast that provenance is easy to lose. A model you cannot reconstruct is a model you cannot defend, and in underwriting, defensibility is the product.
Rule 9: Settle the data terms before you upload a package
Deal packages are confidential, often under NDA, and frequently carry a seller’s financials that are not yours to feed into someone else’s model training. Before the first OM goes up, get three answers in writing: where the data is stored, whether your uploads train shared models, and how you delete your history and export it when you leave.
A small firm handles genuinely sensitive information without an IT department to vet vendors, which makes the contract the only safeguard. “Your data is secure” is a marketing line; a data-processing addendum that commits the vendor not to train on your uploads and to delete on request is an enforceable one. This applies to the general assistants too — know whether the tier you use trains on your inputs before you paste a rent roll into it. If a vendor is vague on any of the three, treat the vagueness as the answer and walk.
Rule 10: Get fluent before you buy a platform
Much of first-pass underwriting is a well-built prompt over ChatGPT, Claude, or Microsoft Copilot plus the model you already run in Excel. An underwriter who can write a clear extraction-and-reconciliation prompt and read the output against these rules will get most of the value with tools the firm already pays for, and will become a far sharper buyer if a platform does turn out to be worth it.
Fluency first does two things. It tells you whether you need a dedicated platform at all — many small firms find the prompt-plus-spreadsheet workflow covers their volume — and it makes you sharp in a demo, because you know exactly what the automation is replacing and where it will cut corners. This is where a short, hands-on session earns out: teach the team to prompt against OMs, rent rolls, and financials, then decide whether purpose-built software is worth the spend. Buy fluency before you buy a platform.
The ten rules as a one-page rulebook
Pin this next to the model and run every AI-assisted underwrite against it.
| # | Rule | The check |
|---|---|---|
| 1 | Screen first | Only underwrite deals that cleared a cheap first cut |
| 2 | Verify extractions | Treat every pulled figure as a claim, not a fact |
| 3 | Ground every figure | Each number links to its source page and line |
| 4 | Reconcile documents | List everything that does not tie out across the package |
| 5 | Own the assumptions | You set exit cap, rent growth, downtime — never the model |
| 6 | Human go/no-go | Capital moves only on an analysis a person interrogated |
| 7 | Use a template | Same fields, same checks, same format on every deal |
| 8 | Keep an audit trail | You can reconstruct what the model touched and what you changed |
| 9 | Settle data terms | Storage, training, and deletion answered in writing first |
| 10 | Fluency before platform | The team can already run the workflow by hand |
Rules 2, 3, and 5 are the ones that move money — a breach there is a bad bid, not a wasted hour. Rules 6 and 9 are the ones that protect the firm. A workflow that honors all ten lets a lean team underwrite more deals, faster, without trading away the judgment that made the firm worth running.
Frequently asked questions
What is AI-assisted underwriting in commercial real estate?
AI-assisted underwriting is the use of a model to handle the mechanical parts of analyzing a deal — reading the offering memorandum, extracting the rent roll and trailing twelve into structured fields, cross-checking the documents, and drafting a first-pass return — while a person owns the judgment: the assumptions and the final decision. The emphasis is on “assisted.” The tool compresses the hours a lean team spends on transcription and reconciliation so the underwriter can spend theirs on whether the deal is real. It is a workflow, not a decision engine, and firms that treat it as the latter inherit its errors.
Can AI underwrite a deal on its own?
No, and building the workflow as if it could is the core mistake. AI can extract numbers, flag discrepancies, and calculate a return on assumptions you give it, but it cannot judge whether an exit cap is defensible, read a seller’s motivation, or decide whether a submarket’s story holds. Those are the parts that set the price and carry the risk. The correct division of labor is AI for the mechanical work and human judgment for the assumptions and the go/no-go. A tool marketed as underwriting a deal end to end is describing a risk you would be buying, not a feature.
What are the biggest risks of using AI in underwriting?
Three. First, a clean-looking wrong number — a figure pulled from the adjacent column of a financial table or a variable extracted incorrectly — that passes review because it looks finished and then moves a bid. Second, letting the model set assumptions like exit cap and rent growth, where it has no genuine market view but full confidence. Third, feeding confidential deal packages into a tool whose data terms you have not checked. Grounding every figure, owning your assumptions, and settling data terms in writing are the rules that address them.
How accurate is AI at reading a rent roll or a T-12?
Modern tools are strong on clean, well-formatted documents and weaker on the scanned, inconsistent, typo-ridden files that fill a real inbox. The accuracy that matters is on your worst documents, not a vendor’s demo. The failure to fear is not a visible error — those get caught — but a plausible figure read off the wrong row or column that looks correct in a tidy table. That is why grounding matters: every extracted number should link to its source page so a person can verify the ones a decision hinges on. Test any tool on your own messy deals before you trust it.
Should AI set the assumptions, like exit cap and rent growth?
No. Assumptions are the levers that determine the bid, and they are pure judgment about a specific market at a specific time. A model will suggest an exit cap with the same confidence it uses for a number it read off the page, but it has no view of the submarket’s next eighteen months and no memory of what traded nearby. Set your own assumptions from your market knowledge and defensible comparables, and let the model calculate the return on them. AI does the math; you own the inputs that matter.
Is it safe to upload confidential deal packages to an AI tool?
Only if the contract says so. Deal packages are confidential and often under NDA, so before uploading, get three answers in writing: where the data is stored, whether your uploads train shared models, and how you delete and export your history. A data-processing addendum that commits the vendor not to train on your data and to delete on request is enforceable; “your data is secure” is not. This applies to general assistants too — know whether the tier you use trains on inputs before you paste a rent roll. A small firm without an IT department has the contract as its main safeguard, so treat vagueness as a reason to walk.
Do I need an underwriting platform, or can I use ChatGPT or Claude and a spreadsheet?
Often a general assistant plus the model you already run covers a small firm’s first pass. A clear extraction-and-reconciliation prompt over ChatGPT, Claude, or Microsoft Copilot will read an OM, pull the numbers, and flag what does not tie out, and at low volume that may be all you need. The limits: general models will confidently misread a figure off a messy T-12 and do not connect to your pipeline, so you verify and manage by hand. Use them to get fluent and to prove whether a platform is worth buying — and never let a model’s figure move money unchecked.
How much does AI-assisted underwriting cost for a small CRE firm?
It depends on whether you buy or build. Off-the-shelf underwriting and deal-management subscriptions for small firms typically run a monthly per-seat or per-firm fee, while a custom-built workflow is a project generally landing in the tens of thousands to low six figures depending on scope. The subscription is rarely the real cost — setup, integration, training, and maintenance make up most of the twelve-month total. Getting fluent with the assistants you already pay for is the cheapest way to find out how much tooling you actually need.
How do I keep an audit trail when AI touches the model?
Preserve the source-linked extractions, note wherever you overrode a figure or an assumption, and keep the reconciliation flags with how you resolved them. The goal is that a lender, an investment committee, or your future self can reconstruct where every number came from and what a person changed. AI produces so much so fast that provenance is easy to lose, and a model you cannot reconstruct is one you cannot defend — in underwriting, defensibility is the product.
How do I get my team fluent enough to underwrite with AI safely?
Start with the tools you already own and a small set of real deals. Teach the team to write a consistent extraction-and-reconciliation prompt, to demand source-linked figures, to fill assumptions themselves, and to read every output against a rulebook like the one above. A short, hands-on session focused on prompting for OMs, rent rolls, and financials builds the judgment that keeps AI’s speed from becoming a liability, and it makes the buy decision honest — many firms find the prompt-plus-spreadsheet workflow covers their volume.
Where to start
The first question is not which underwriting tool to buy. It is whether your team can already run a disciplined first pass by hand — extract, ground every figure, reconcile the documents, set the assumptions themselves, and keep the decision human. A firm that can do that with the tools it already owns will know exactly what a platform would need to add before it is worth the spend. A free AI-readiness assessment gives you that read: a short working session that looks at your deal flow, your current underwriting process, and where your analyst hours actually go, then returns an honest recommendation on the right next step. Book a free AI-readiness assessment before you commit to software your process may not need.
Arthur Wandzel
