The acquisition model for a 48-unit apartment building looks complete. Rent roll imported. Historical expenses categorized. Debt assumptions entered. Ten-year cash flow calculated.
The projected return clears the investment team’s threshold.
One line is wrong.
A building-services agreement charges $4,200 each month. The extraction system captures the amount correctly but classifies it as an annual expense. The model therefore includes $4,200 instead of $50,400.
Nothing breaks. The spreadsheet balances. Every formula works. Net operating income is overstated by $46,200 a year, and at a 5.5% capitalization rate, that single classification error inflates the implied value by $840,000.
Automation has not created the mistake. It has removed the friction that might have exposed it — and carried the result through every downstream calculation in seconds.
The property and figures in this opening are illustrative.
The example is deliberately institutional: a $4,200 monthly service contract is plausible at that scale and makes the error easy to see. But the same evidence-to-assumption failures affect smaller investors, often with fewer people available to catch them. Underwriting brings the series’ central problem into an explicitly financial model.
Underwriting is not a spreadsheet
The spreadsheet is where an investment thesis becomes numerical. It is not where that thesis begins.
An acquisition underwrite usually combines several different jobs:
- Collect evidence from rent rolls, leases, operating statements, service agreements, tax records, insurance information, inspections and market data.
- Normalize that evidence into comparable units and categories.
- Decide which historical figures are recurring, which are exceptional and which should change under new ownership.
- Form assumptions about rent, vacancy, operating costs, capital expenditure, financing and exit conditions.
- Test scenarios and decide whether the expected return compensates for the risk.
Only the last part looks like a decision. The decision is already being shaped in every step before it.
If a monthly charge becomes annual, if square feet become square meters, if a rent concession is treated as contractual rent or if a one-time repair is projected forever, the model can remain internally consistent while becoming economically wrong.
That is why underwriting should be designed as an evidence chain rather than a completed template.
AI is well suited to the mechanical layer
Real-estate underwriting still depends heavily on unstructured and inconsistently formatted information. Leases describe similar obligations differently. Operating statements use different account names. Rent rolls arrive in spreadsheets, PDFs and exports with different definitions. Important terms can sit inside schedules, amendments or scanned documents.
This is exactly where AI can remove low-value friction.
It can extract lease dates, rents, escalations and options; map differently named expenses into a common chart of accounts; identify missing documents; compare a rent roll with lease abstracts; flag values outside expected ranges; and prepare alternative scenarios from assumptions the investment team defines.
The opportunity here is credible for an unglamorous reason: the work is large, repetitive and information-heavy, and much of it can be completed without asking the system to decide what will happen next.
But these tasks require different methods. Document extraction is not the same as forecasting. Anomaly detection is not investment judgment. Scenario calculation should come from a controlled financial model, not a language model improvising arithmetic. A system can coordinate these capabilities, but it should not blur their responsibilities.
That division does not diminish the role of AI. It directs AI toward the work where scale is useful and makes the consequential decisions easier to inspect.
Speed becomes dangerous at the assumption boundary
Extraction asks, “What does the document say?” Underwriting asks, “What should we believe will happen?”
The two questions meet at the assumption boundary.
A trailing operating statement may accurately report last year’s property tax, but an acquisition could trigger reassessment. Current insurance cost may be factual, while renewal cost remains uncertain. In-place rent is known; achievable market rent is a thesis. A recent repair is documented; whether it will recur is a judgment. An exit capitalization rate is not extracted from any property document at all.
What the evidence says
- Last year’s property tax
- Current insurance premium
- In-place contractual rent
- A recent repair expense
- No property document specifies an exit cap
What the underwrite must assume
- Tax after acquisition or reassessment
- Expected renewal premium
- Achievable market rent
- Whether the repair will recur
- An appropriate exit capitalization rate
AI can help research and test these assumptions. It cannot convert them into facts by placing them in a polished model.
Five errors become especially dangerous when automated:
The faster the model updates, the more easily one weak assumption can propagate. Speed is beneficial only when the system also makes uncertainty and dependence visible.
Design the underwrite to invite challenge
Traditional model review often begins with the outputs: purchase price, yield, debt coverage, cash-on-cash return and IRR. An AI-supported workflow should make it easier to challenge the inputs before debating the outputs.
That requires several controls.
Preserve provenance at the field level
Each material fact should link to its source, page or cell, extraction time and transformation history. A reviewer should be able to click an expense assumption and see the document language that produced it.
Provenance does not prove the assumption is right. It makes the reasoning inspectable.
Separate facts from assumptions
The last twelve months of insurance expense is a historical fact when supported by the records. Next year’s premium is an assumption. The model should label them differently, preserve both and show who approved the forward figure.
Show ranges before confidence hardens
Base, downside and upside cases should not be three arbitrary percentage changes applied to the same optimistic thesis. They should expose the assumptions that drive the outcome: rent growth, vacancy, expense inflation, capital work, financing and exit pricing.
Sensitivity analysis is most useful when it identifies what would have to be true — not merely what the return becomes.
Route exceptions, not every field
If AI extracts 500 fields, asking a professional to recheck all 500 recreates the original workload. The system should concentrate review on low-confidence extraction, conflicting sources, unusual units, missing periods and assumptions with material effect on value.
Human review has to be selective, but the selection logic must itself be tested. Random sampling — not only reviewing cases the system has flagged — is one way to test the routing logic, because a rule cannot expose its blind spots using only the cases it selects. A field should not avoid review merely because the model sounds confident.
Low confidence, conflict, unusual unit or material effect
Tests the cases the routing logic did not select
Retain the override and its reason
When an investment professional changes an extracted value or forecast assumption, the system should preserve the original, the revision, the reviewer and the reason. Overrides are not evidence that the system failed. They are evidence about where professional judgment enters — and data for improving the workflow.
Professional judgment is a control, not a ceremony
“Human in the loop” can mean little more than placing an approval button after an automated recommendation.
Meaningful review requires three things: the evidence needed to challenge the output, the authority to change or reject it and enough time to do so before the decision is made.
The Royal Institution of Chartered Surveyors’ professional standard on responsible AI use, effective March 9, 2026, makes this concrete for RICS members and regulated firms. For AI outputs that materially affect a surveying service, it requires professional judgment about reliability, documentation of relevant assumptions and concerns, and responsibility by an appropriately qualified, named surveyor. For automated or high-volume outputs, it accepts that scrutinising every one is “neither necessary nor proportionate” — and then requires “randomised dip samples of the outputs at regular intervals” instead, with firms remaining accountable for each output. Scale changes the method of review, not the fact of it.
The exact professional and regulatory duties vary by role and jurisdiction. The broader operating principle does not: the person responsible for the investment recommendation must be able to see how the system reached it and must remain able to disagree.
AI should prepare the case and pressure-test it. It should not become the unnamed author of the investment thesis.
Measure whether the workflow improved — not how fast the model filled
Time saved is useful, but it is an incomplete measure. A system that reduces initial model preparation by four hours and adds three hours of correction has saved one hour, not four. A system that accelerates screening while increasing missed-risk events may have made the investment process less efficient in the only sense that matters.
A serious implementation should track:
- Review time after automated preparation
- Extraction error by field type and source format
- Assumption override rate and reason
- Exceptions caught before approval
- Forecast-to-actual variance in rent, expenses and capital work
- Risks missed and discovered after acquisition
- Decisions declined or repriced because new evidence changed the thesis
These measures do more than evaluate the tool. They reveal where the underwriting process itself relies on weak data, inconsistent definitions or judgment that has never been made explicit.
NIST’s AI Risk Management Framework organizes AI risk work around four functions: Govern, Map, Measure and Manage. Applied to underwriting, that means assigning decision rights, understanding the specific use and its consequences, testing the system against relevant failure modes and improving controls as actual outcomes arrive.
The feedback loop should not train a system blindly to imitate past approvals. It should help the organization understand which assumptions were repeatedly wrong, which exceptions mattered and which evidence changed decisions.
What should have caught the $4,200 line
The $4,200 figure in the opening was extracted correctly. The failure occurred when frequency was normalized.
A dependable workflow would preserve the source language, classify the billing frequency with a confidence score and compare the assumption with the series of payments in the operating history. The mismatch between one annual charge and twelve prior payments would enter an exception queue. The reviewer would see the $46,200 annual difference and its $840,000 effect at the selected capitalization rate before approving the case. In this illustrative case, the corrected return no longer clears the investment team’s threshold. The acquisition goes back for repricing rather than being approved at the original price.
$840,000 value effect
The control is not “have a human check the AI.” It is more specific:
Show the reviewer the assumption most capable of changing the decision, the evidence behind it and the consequence of being wrong.
The model still updates in seconds. The investment team still saves the work of manually populating hundreds of fields. But speed is now paired with traceability and challenge.
That is the useful future of the AI underwriter. Not a machine that produces a finished investment answer, but a system that makes the evidence easier to assemble, the assumptions harder to hide and the decision more accountable.
Sources
- Royal Institution of Chartered Surveyors, Responsible use of artificial intelligence in surveying practice, first edition, September 2025; effective March 9, 2026.
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023.
Reactions
No reaction from you
