How to use AI for real estate underwriting without trusting the wrong answer

The most dangerous AI property analysis is not obviously absurd. It looks polished. The address is nearly right, the numbers add up, and the explanation sounds confident. One comparable is in a stronger town. The square footage came from the wrong record. The repair allowance assumes the roof is fine. The output can be internally neat and financially wrong.

Responsible AI underwriting begins with a simple role: the system is a research and calculation assistant. A person remains responsible for verifying the subject, evaluating evidence, inspecting condition, choosing assumptions, understanding risk, and approving any action. This guide is educational, not an appraisal, inspection, lending decision, legal opinion, tax advice, or guarantee of investment results.

1. Give AI a useful job, not total authority

AI is good at transforming scattered information into a structured draft. It can extract listing facts, identify missing inputs, search for public sources, propose comparable candidates, summarize inspection notes, categorize repairs, and calculate several exit scenarios. Those tasks reduce clerical work and make gaps easier to see.

It is not physically present. It cannot smell moisture, open an electrical panel, confirm whether a bedroom is legal, know what a contractor saw behind a wall, or promise a buyer will pay the modeled value. It does not know your liquidity, lender relationship, experience, or tolerance for a delayed project unless you provide that context.

Use three output labels

• Observed: directly supported by a listing, public record, photo, inspection, bid, or other identified source.
• Calculated: math derived from stated inputs, such as projected profit or monthly cash flow.
• Inferred: a reasonable but unverified interpretation, such as likely condition from limited photos.

If a tool blends those categories into one confident narrative, the user cannot tell what to verify. NIST's generative AI risk profile describes confidently presented erroneous content as confabulation and notes the risk when people act on it in consequential decisions. The practical defense is evidence, labels, testing, and human review.

Scenario: Erin changes the assignment. Erin first asks for “the correct offer” on a dated bungalow. The output looks precise, but it cannot know her financing or required return. She reframes the request: extract facts, identify unknowns, research comparable candidates, and calculate three offers from her stated repair and profit assumptions. The new output supports a decision instead of pretending to make one.

2. Verify the subject property before analyzing it

A correct calculation tied to the wrong house is still wrong. Begin with the full street address, city, state, ZIP code, unit, property type, legal use, beds, baths, living area, lot, year built, and listing source. Compare listing fields with assessor, permit, deed, tax, or other appropriate local records. Record disagreements instead of silently choosing the most favorable number.

Repeated street names create subtle failures. A comparable sale may be real and recent but located across a municipal boundary with different taxes, schools, services, zoning, or buyer demand. A unit number can disappear from a listing URL. Finished basement area can be presented as though it were above-grade living area. A two-family can be marketed like a large single-family home.

Stop conditions for the subject

• The city or unit is uncertain.
• Legal use conflicts with the marketing description.
• Square-footage sources materially disagree.
• Bedroom or bathroom counts do not match plans, records, or visible condition.
• The listing URL resolves to a different property.
• The property is unusual enough that ordinary comparison methods are weak.

Do not reward uncertainty with a wider prompt. Resolve it through direct evidence. Use property research to keep the source and the fact together, and follow the subject verification steps in the first-property analysis guide.

3. Never accept comparable sales just because they are nearby

Comparable selection is a judgment process, not a distance sort. Begin with closed sales that compete with the subject in the relevant market. Review location, sale date, property type, size, room count, lot, condition, renovation quality, functional utility, parking, view, concessions, and unusual circumstances.

Proximity helps, but municipal lines, neighborhood boundaries, school assignments, flood exposure, traffic, and development patterns can change buyer behavior over a short distance. If the best available comp differs from the subject, explain the difference and how the market reacts to it. Do not invent a dollar adjustment because a feature feels more valuable.

Fannie Mae's appraisal guidance says adjustments should reflect market reaction and that no two properties or transactions are usually identical. An investor's analysis is not an appraisal, but the underlying discipline is useful: choose relevant evidence, analyze differences, and explain the conclusion.

Reject attractive but misleading comps

A fully renovated 4,500-square-foot luxury property should not establish value for a 1,700-square-foot ranch merely because both are on the same road. The cheapest distressed sale should not automatically become the finished-value ceiling. Active listings show competition, not closed proof. Pending prices may be unavailable or change before closing.

Scenario: Caleb catches the town-line error. Caleb analyzes a 1,850-square-foot home in Holbrook. One suggested comp is two miles away in a higher-priced town and another is a 3,900-square-foot custom property. He removes both, adds two same-town sales with similar utility, and lowers the value range. The correction reduces his maximum offer by $41,000 before he signs a contract.

4. Require a value range with evidence, not one magic ARV

After-repair value is an opinion about a future market condition after a defined renovation. It depends on what the finished property will actually be, when it will be sold, and how buyers respond. Present a supported range, a working point estimate, and the reasons the result could move.

Every candidate comp should display its address, status, sale date, price, source, property characteristics, distance or market relationship, and why it was included. If the source is unavailable, mark it for verification. If tax and listing records disagree, show both. A polished number without inspectable evidence is not more reliable because it has decimal places.

The federal automated valuation model rule applies to specified mortgage and securitization uses, not an ordinary investor's private deal screen. Still, its quality-control themes are instructive: confidence in estimates, protection against data manipulation, conflict controls, testing, review, and compliance with nondiscrimination laws. An investor tool should not claim regulatory compliance merely because it follows similar principles.

Use deal analysis to test the point estimate and downside values. If a deal only works at the top of the range, that fragility belongs in the decision.

5. Never let photos become a complete repair budget

Images can help identify visible conditions and organize a preliminary scope. They cannot reveal every concealed system, establish code compliance, diagnose structural movement, test environmental hazards, or price local labor reliably. A photo-based result should say what is visible, what is uncertain, and which specialist should verify the condition.

Turn observations into scope questions

“Old bathroom” is not a budget line. Record the location, observed condition, proposed action, quantity, finish standard, dependencies, and uncertainty. Separate visible work from allowances and unknowns. Then obtain contractor, inspector, engineer, environmental, permit, or supplier input appropriate to the property.

Scenario: Nia refuses the photo estimate. Nia receives a $58,000 preliminary scope from listing photos. During access, her electrician identifies obsolete service equipment, and a roofer finds multiple layers over damaged decking. She replaces the photo allowance with written scope and bids, raising the budget to $91,000. Because the AI draft clearly labeled the systems as unverified, she had not used the lower number as a promise.

Build the detailed scope with the beginner renovation budget guide and maintain approvals through rehab estimating.

6. Separate source facts, assumptions, and deal math

AI can calculate perfectly from weak assumptions. Keep three layers visible so a reviewer can trace the answer.

1. Evidence layer: listing facts, public records, comparable sales, photos, reports, bids, lender terms, and source dates.
2. Assumption layer: working ARV, scope, contingency, schedule, interest rate, selling costs, vacancy, rent, and required return.
3. Calculation layer: acquisition cash, financing cost, holding cost, projected profit, return, rental cash flow, refinance proceeds, and maximum price.

If an assumption changes, the calculation should update without rewriting the underlying evidence. If new evidence arrives, record which assumption it supports or replaces. This preserves a decision trail and prevents a persuasive generated explanation from becoming the source of its own facts.

Quick rules remain screens. The 70% rule cannot know your financing, selling cost, timeline, tax position, or required return. AI should explain that limitation and show the full cost stack rather than hiding behind a familiar formula.

7. Treat confidence as a description of evidence completeness

A confidence score is useful only when you know what creates it. It should rise when the address is locked, core facts agree, relevant comps are verified, condition evidence improves, bids replace allowances, and financing terms become specific. It should fall when facts conflict, sources are sparse, the property is unusual, or the analysis depends on distant or dissimilar evidence.

Confidence does not mean probability of profit. A well-documented deal can still lose money. It means the analysis is more complete and traceable. Ask the system to list missing items and show how each could affect the result.

• Low confidence: use for lead triage and fact collection, not an offer.
• Moderate confidence: use to plan diligence and negotiate within a protected range.
• Higher confidence: use after important property, scope, value, financing, and exit inputs have direct support, while preserving downside analysis.

Do not let a provider name, advanced terminology, or long explanation substitute for evidence quality. The standard applies regardless of which system serves the request.

8. Never let AI silently write to money fields

An AI output should arrive as a draft. ARV, repair budget, offer ceiling, rent, schedule, and financing assumptions should change only after a person reviews the evidence and accepts the proposal. The system should preserve the prior value, proposed value, reason, source, reviewer, and time.

Human approval is not a ceremonial button. The reviewer needs enough context to notice a wrong address, weak comp, missing system, stale rate, or impossible timeline. Require additional approval when a change crosses a material threshold or affects a partner, lender, tenant, buyer, or regulated decision.

The CFPB has stated that creditors using complex algorithms still must provide specific and accurate reasons for adverse credit actions. Rehabfolio is an investor project tool, not a creditor making consumer approval decisions, but the broader lesson is sound: “the system decided” is not an explanation. People need understandable reasons tied to the factors actually used.

Scenario: Mateo preserves the audit trail. Mateo's revised comp set raises working ARV from $610,000 to $638,000. Instead of overwriting the field automatically, the draft shows the three new closed sales, rejected outlier, and downside range. Mateo and his partner approve $625,000 as the working case and keep $600,000 as the stress case. The record shows what changed and why.

9. Use this repeatable AI underwriting workflow

1. Define the question. State flip, rental, refinance, wholesale, or construction goal and the decision you need to make.
2. Provide complete input. Add the full address, listing source, facts, photos, known condition, financing, timeline, and return requirement.
3. Lock the subject. Resolve city, unit, legal use, property type, room count, and square footage conflicts.
4. Collect evidence. Gather public records, relevant closed sales, rental evidence, inspection findings, bids, and loan terms.
5. Review exclusions. Inspect rejected comps and missing facts, not only the evidence the system selected.
6. Build the scope. Convert visible conditions into line items, allowances, specialist questions, and contingency.
7. Run likely and downside cases. Change value, repairs, timing, financing, rent, vacancy, and selling costs.
8. Approve deliberately. Accept, edit, or reject each material draft and preserve the reason.
9. Refresh when evidence changes. Update the analysis after inspection, bids, appraisal, lender terms, or market changes.

Use reports and collaboration to show partners the assumptions and evidence, then compare a completed project's outcomes with the sell, rent, or refinance guide. The objective is not to remove judgment. It is to make judgment faster, more consistent, and easier to audit.

Frequently asked questions