A model can produce a number. It cannot be responsible for one.
That distinction is not a technical limitation waiting to be solved. It is the shape of the problem.
Property valuation has genuine, useful machine learning in it. It also has a layer that resists automation for reasons that have nothing to do with compute.
Where models genuinely help
Models are strong wherever the input is structured, plentiful and checkable. They weaken wherever it is sparse, unrecorded or contested.
Finding candidate comparable sales
Sifting thousands of recent transactions for plausible matches is exactly the kind of search a machine should do. It saves an appraiser real hours.
Detecting inconsistency
Flagging a report where the sketch, the room count and the stated living area disagree. Rule checking at scale is a genuine win.
Reading unstructured records
Pulling permits, deeds and prior listing history into structured fields. Tedious, mechanical, well suited to software.
Estimating at portfolio scale
For a lender reviewing thousands of properties, a model that is roughly right in aggregate is genuinely useful, even where any single estimate might be off.
Condition it has never seen
Two identical houses on the same street can be decades apart in condition. Records cannot separate them. Nothing in the training data will.
Thin and unusual markets
Acreage, custom builds, mixed use, anything rare. Few genuine comparable sales means little to learn from, and confidence drops exactly where stakes are often highest.
Explaining itself
An appraisal has to show why the market supports each adjustment. A number without a defensible chain of reasoning does not meet that bar.
Carrying the consequence
Someone has to be accountable when a value is wrong. Licensing, professional standards and liability attach to a person. They do not attach to a model.
The line is accountability, not accuracy
It is tempting to frame this as a race, where models improve until they overtake appraisers. That misreads the problem. Even a model that was more accurate on average would still not hold a licence, could not be disciplined by a state board, and could not stand behind a value with supported reasoning.
Lending rests on somebody being answerable. That is a structural requirement, not a gap in the technology.
The version that actually works
The useful arrangement is not one replacing the other. It is the machine doing retrieval, extraction and checking, and the appraiser doing selection, adjustment and reconciliation.
Machine proposes, appraiser disposes. Software surfaces forty candidate sales. The appraiser picks six and explains why those six. The search was mechanical, the choice was not.
That split already describes most serious appraisal technology. What it does not describe is a free estimate on a consumer website, which is a model on its own with nobody standing behind it.
The accountable layer
Three licensed practices doing the part that does not hand over.
Tyler Kane of TyCorp Inc, Certified Residential Appraiser and FHA approved, north Phoenix and the northeast Valley.
TyCorp Inc, Arizona appraisalsWeston Groll of Apex Appraising, Utah Certified Residential Appraiser covering 52 cities across 8 counties.
Apex Appraising, UtahPocket Appraisal, for residential valuation work anywhere in the country.
Pocket AppraisalCommon questions
Can AI do a home appraisal?
Are appraisers using AI already?
Why can a model not just learn condition?
Will lenders stop requiring appraisals?
Building something in this space?
Tell us which part of the valuation problem you are working on.