An automated valuation looks brilliant on a house that's for sale and falls apart on one that isn't. That gap is the whole problem, and it's the one almost every team underestimates.
Here is the simple version. An automated valuation model, an AVM, takes a property and predicts what it's worth, the way Zillow's Zestimate or Redfin's estimate do. We have built one as the core of a real-estate platform, an engine that suggests a property's rent or sale value, and the hard part was never wiring up a model. It was making the number trustworthy enough that someone would act on it. Because the moment a buyer, an investor, or a lender makes a decision on your estimate, a wrong number isn't a bad demo, it's a real loss attached to your name.
The on-market illusion
Look at the published accuracy of the two best consumer models and the trap jumps out. Zillow's Zestimate runs about 1.8% median error on homes that are actively listed and around 7% on homes that aren't. Redfin's estimate is nearly identical, tight on-market and roughly four times worse off-market. Same pattern, same factor, every time.
The on-market number is flattering for a reason that has nothing to do with the model being smart. When a house is listed, the model gets to see the agent's list price and a stream of fresh signals, so it's half-grading its own homework.
Off-market, it has to reason from stale public records and comparable sales, and it can't see that the kitchen was just renovated or that the roof is failing. That's where real valuation lives, and it's where accuracy collapses. If your platform shows a confident number on a house nobody has listed, you are standing on the shakier three-quarters of the problem, and that is exactly the number people will lean on hardest.
The moat was never the model
It is tempting to think the differentiator is a cleverer machine-learning model. It isn't, and treating it that way is how teams burn months. The model is the commodity. The moat is three unglamorous things underneath it: comp selection, calibration, and data coverage.
Comp selection is the real intelligence. Picking the right handful of comparable sales, adjusting for the differences that matter, and knowing when there simply aren't enough recent comps to say anything honest, that is what an experienced appraiser does in their head and what your engine has to do at scale.
Calibration is making sure a number your model is confident about is actually as reliable as it claims. And coverage is the boring grind of having clean, current data on enough properties and transactions that the comps exist at all. Institutional, credit-grade valuation models are judged on metrics like how often they land within ten percent of the truth, and the serious ones treat that bar as the product, not the model architecture.
The AI angle, done grounded
There is a real AI angle here, and it's worth building, but only the disciplined version. The useful move is to feed the model signals a flat statistical AVM never had: condition and quality read from listing photos, renovation history, hyper-local features that move price on one street and not the next. That genuinely improves the estimate when it's grounded in real evidence about the actual property.
The version that hurts you is the one that uses AI to project false confidence. A valuation that comes back as a single, precise-looking dollar figure on a property the model can barely see is lying with a straight face. The honest design shows a number with a confidence range that's tight when the comps are strong and wide when they're thin, and that widens automatically off-market. That's not a weaker product.
A single precise-looking figure on a property the model can barely see is lying with a straight face. A buyer or a lender trusts a range they can reason about far more than a single number that turns out wrong, and the range is what keeps you out of trouble when it does.
Valuation is now a compliance surface too
One more thing the generic build advice skips: as of late 2025 a federal rule on automated-valuation quality control took effect, and it folds a nondiscrimination requirement into the controls you're expected to have. The broader fair-lending backdrop is in motion, with the disparate-impact framework reinstated in 2023 and a rollback proposed as of January 2026, so you can't hardcode to one regime. The point for a builder is concrete: a valuation model can produce discriminatory outcomes without anyone intending it, and "the algorithm decided" is not a defense. The quality controls, the bias testing, and the audit trail are part of the engine now, not a compliance afterthought. We thread that through the compliance piece.
What's still standing in 2028
By 2028 anyone can call a pretrained model and get a number. The number itself is worth nothing. What compounds is everything that makes it trustworthy: your comp logic, your calibration, the proprietary transaction and outcome data only your platform sees, and the audit trail that proves the estimate was produced fairly. That is the AVM moat, and it's the opposite of the part that demos well. The teams still standing are the ones who treated the valuation as a number they'd have to defend, not a feature they shipped.
FAQ
Why is an automated valuation accurate on listed homes and not on unlisted ones? Because on a listed home the model can see the agent's list price and a stream of fresh signals, so it is half-grading its own homework. Off-market it reasons from stale public records and cannot see that the kitchen was renovated or the roof is failing.
How accurate is the Zestimate? Zillow publishes roughly 1.8% median error on homes actively listed and around 7% on homes that are not. Redfin's estimate is nearly identical, tight on-market and roughly four times worse off-market. The same pattern and the same factor show up every time.
What actually differentiates an AVM if the model is a commodity? Three unglamorous things underneath it: comp selection, calibration, and data coverage. Picking the right comparable sales, adjusting for the differences that matter, and knowing when there are too few recent comps to say anything honest is the real intelligence.
Should a valuation show a single number or a range? A range. A single precise-looking figure on a property the model can barely see is lying with a straight face. An honest design tightens the band when comps are strong and widens it when they are thin, and widens automatically off-market.
Is a valuation model a compliance surface? Yes. A federal rule on automated-valuation quality control took effect in late 2025 and folds a nondiscrimination requirement into the controls you are expected to have. The wider fair-lending backdrop is in motion, so you cannot hardcode to one regime.
How should a team measure whether its AVM is trustworthy? On off-market accuracy, not the flattering on-market number. Pull properties that recently sold but were not listed on your platform when you valued them, and compare the estimate to the real sale price. That error is what users are actually trusting.
What 2muchcoffee covers
We build production valuation engines and the grounding around them, the comp selection, the calibration, the confidence modeling, and the bias and audit controls that let a platform stand behind a number. We ship domain-specific AI to production, not generic AI consulting. If you're putting a valuation in front of buyers, investors, or lenders and you can't yet defend the off-market number, that's the conversation to have before it's live. The plain way in is the AI work we do.
One concrete action
Measure your own off-market accuracy, not your on-market accuracy. Pull a set of properties that recently sold but were not listed on your platform when you valued them, and compare your estimate to the real sale price. The error you find there, not the flattering on-market number, is what your users are actually trusting. If it's wide and you're still showing a single confident figure, the fix isn't a better model, it's an honest confidence band. This is one layer of building a real estate platform, and it's the layer where trust is won or quietly lost.
