Paste a listing: what good AI underwriting should never do

The easy part and the hard part

Pasting a listing into an AI and getting a number back is the easy part — any chatbot will happily oblige. The hard part is that the number is attached to real money: your offer, your lender's capital, your margin. So here is the standard we hold our own underwrite to, written as a list of things good AI underwriting should never do. If your current tool does any of these, distrust it.

1. Never pull comps from the wrong city

Street names repeat across neighboring towns, and models love to "helpfully" grab a comp from the wrong one. A sale from a pricier neighboring city can inflate an ARV by six figures in the wrong direction. This is the classic hallucination: the comp is real, the city is wrong. The guard is a hard city-lock — comps and tax data must come from the subject's city, full stop. And when a listing sits far outside the market's normal band — an ultra-luxury property with zero same-city comps — the honest move is to drop it and say so, not to invent a valuation.

2. Never mix mansion square footage into a ranch comp set

A 4,500 sq ft sale is not evidence for a 1,900 sq ft ranch, even on the same street. Size-band mixups are the second classic failure: the model averages everything nearby and the ARV lands somewhere no actual buyer would pay. Comps need to be filtered to the subject's size and property type, and luxury outliers in an ordinary neighborhood need to be excluded, not averaged in.

3. Never hand you a number with no evidence

An ARV without comps is a vibe. Good AI underwriting shows its work: which sold comps it used, where each one came from, and the public-record context around the subject. That is why every Rehabfolio Web CMA ships with an evidence panel — the sold comps behind the estimate, clickable source links, and tax/assessor data (assessed value, annual tax, parcel, living area, year built). Thirty seconds of scanning the panel tells you more than thirty minutes of re-prompting a black box.

4. Never floor ARV to the cheapest comp

This one wears a conservative costume. "Just use the lowest comp" feels safe, but the cheapest closed sale is usually the most distressed sale — the estate liquidation, the deferred-maintenance special. Floor every ARV to it and you will chronically under-bid, lose good deals, and never know why. The ARV should be the research memo's point estimate, defended by adjusted comps — with the cheap comp explained, not obeyed.

5. Never write to your money fields

An AI that silently updates your ARV, repair budget, or max offer is not an assistant — it is a liability with a keyboard. Every AI output should arrive as a draft in a review panel, and nothing should persist until a human clicks accept. In Rehabfolio this is a hard rule: underwrites, scopes, daily priorities — all drafts, always. You can edit, accept, or discard. The audit log records what ran and which provider served it.

When to distrust the model (even ours)

Sparse listings. "Great opportunity — bring your contractor!" with no square footage, no bed/bath count, and three dark photos is not an underwrite waiting to happen; it is a guess waiting to happen. Watch the confidence signal: when core facts are missing, the draft tells you what to add to raise confidence instead of pretending it knows. If the listing contradicts itself — a "fully renovated" remark next to a 1978 kitchen photo — treat the output as a starting bid for your own research, not an answer.

What to paste for better results

The model is only as good as the facts you feed it. Paste the full MLS remarks, list price, beds, baths, square footage, and year built — and include the Zillow or Redfin link. Rehabfolio parses the subject address straight out of the URL slug, which anchors the city-lock and the tax-record lookup before the research even starts. More facts in, higher confidence out.

AI underwriting is worth trusting when it shows its evidence, respects its own uncertainty, and waits for your accept. Anything less is a random-number generator with good typography.