Every estimated price on a property website comes from a model, a person or both. Models do some jobs well and others badly. Knowing which is which lets you use their output without trusting it too far.
What a pricing model does
People call a pricing model AI. In practice it is a statistical method. It takes many past sales, with details such as the tower, its size, the floor level and the sale date, and works out how those details relate to the price. To estimate a new price, it applies the relations it found. The result is a likely price, ideally with a range around it. The model reads numbers, not sentences, and it knows only what its data contains.
Ask it about a two-bedroom flat on a middle floor in a busy building and it has many close matches. Ask it about a one-off penthouse and it has almost none. That difference explains most of what follows.
Where models do well
Three strengths stand out. The first is scale. A person can study a few dozen sales, while a model can study millions. dataHabibi, an AI-powered Dubai housing data platform, reads more than 12 million transactions registered with the Dubai Land Department. At that size, patterns appear that no person would spot in a list, such as how price per square foot shifts with floor height or unit size.
The second is speed. A model can refresh every building each day, which no analyst can do by hand, and that suits a market where new sales register all the time.
The third is consistency. The model applies the same method to every building, with no mood and no sales target. Two flats with the same features get the same treatment, whoever asks.
Where models do badly
Unusual units come first. A rare penthouse, a flat with an odd layout or a building with few sales gives the model few comparisons, so its estimate is weak.
Sudden shocks come second. A model learns from the past, so a policy change, a major launch or a sharp swing in demand appears in its output only after sales record it. A model that learned mostly from a period of steady growth has little to say about a sudden fall. It can show that a price moved, but it rarely says why.
Missing information comes third. A model has no view of how well a flat is kept, what a tenant’s lease says or how loud the road outside is, yet these change what a buyer will pay.
Treat the output as one input
Use the model’s figure as one of several, next to the asking price and what you see at a viewing. A simple rule helps: if the figure falls outside the range of recent sales for near-identical flats in the same building, treat it as a question to investigate, not an answer.
Take a case made up for this article: a model puts a flat at AED 1,200,000 while the last three sales of the same layout in the building fall between AED 1,300,000 and AED 1,350,000. The gap deserves a call to the agent. Perhaps the flat needs repairs. Perhaps the model missed a recent rise. Either way, you learn something. For a large purchase, a human valuer adds a view that no model can give.
What good tools disclose
A trustworthy tool tells you where its data comes from and when it last updated. It shows how many sales stand behind a figure, gives a range instead of a single number and says what the model leaves out. dataHabibi prices each building from sales that happened, not from asking prices, and adds a forecast. Ask also whether the tool reports its past misses. One that shows only its hits is harder to trust. Be wary of any tool that offers a precise figure with no source, no date and no range.
A model reads patterns across many sales well and judges one odd flat poorly. Use it for the first job and check the second yourself. For sold-price history by building, visit datahabibi.ae.
