I’ve been comparing AI-generated property valuations against actual sale prices for the past six months. The results are interesting, sometimes accurate, often wildly off, and always missing context that actually matters to buyers.

Most automated valuation models (AVMs) pull data from public records, recent sales, and property characteristics. They’ll tell you a three-bedroom house in Epping with a 600-square-meter block should be worth between $1.4 million and $1.6 million. That’s a $200,000 range, which isn’t particularly useful if you’re trying to set a reserve price or make an offer.

The bigger problem is that these models don’t account for what’s actually happening on the ground. They can’t see that the house backs onto a busy road, or that the owners renovated the kitchen last year with high-end appliances, or that there’s a proposed development next door that’ll block the afternoon sun.

Where AI Valuations Work

For broad market analysis, the models are decent. If you want to understand general price trends across a suburb or identify which areas are appreciating faster, AI tools can process thousands of sales in minutes and show you patterns.

I’ve used them for client presentations to demonstrate how a particular street’s median price has changed over five years. The visualizations are clean, the data’s usually accurate at that aggregate level, and it saves me hours of manual research.

They’re also useful for flagging properties that might be over or underpriced relative to comparable sales. An AI model can quickly tell you that a property listed for $2.2 million is significantly above recent sales of similar homes in the area. That’s valuable information, even if it doesn’t tell you whether the premium is justified.

Some lenders are using AVMs for preliminary loan assessments. If you’re refinancing and you just need a ballpark valuation, an AI model can provide that instantly rather than waiting days for a human appraiser. For routine transactions where precision isn’t critical, that efficiency makes sense.

Where They Fall Apart

Individual property valuations for unique homes are where AI models struggle. I saw a heritage-listed terrace in Paddington valued by an AVM at $1.8 million. It sold for $2.4 million. The model couldn’t account for the meticulous restoration work, the rarity of period features in original condition, or the fact that the house had been featured in architecture magazines.

Location nuances are another gap. Two houses on the same street, same size, same age can have vastly different values based on which side of the street they’re on, proximity to parks, views, noise levels. AI models use broad location data but they’re not walking the street to see what it actually feels like.

I’ve seen valuations completely miss development potential. A rundown house in Chatswood was valued at $1.3 million based on the existing structure. It sold for $2.1 million because buyers recognized it was on a corner block in an R3 zone where they could build a duplex. The AI saw an old house; buyers saw future value.

Renovation quality is almost impossible for algorithms to assess from data alone. An AI model might see “renovated kitchen” in the listing description and apply a standard uplift. But there’s a massive difference between an IKEA kitchen installed by the owner and a custom Caesarstone and Miele setup done by a professional. That difference can be $100,000 or more in final sale price.

The Data Quality Problem

These models are only as good as the data they’re trained on. In Sydney, property data’s fragmented. Sales prices are public, but renovation histories, structural issues, easements, and planning approvals aren’t always captured in the datasets that AVMs use.

I’ve noticed AI valuations are more accurate in newer suburbs with homogeneous housing stock. In Kellyville or Marsden Park where estates have similar designs and recent sales data is abundant, the models do reasonably well. In older, more diverse suburbs like Balmain or Newtown, accuracy drops significantly.

Partial data creates weird outputs. One AVM valued a property based on land size but missed that half the block had a steep slope that made it unbuildable. Another valued a unit without accounting for its location on a ground floor next to the garbage room, which significantly affects buyer appeal.

The Human Element

Buyers don’t make decisions purely on comparable sales data. They’re influenced by emotion, future plans, and personal circumstances. A family desperate to get into a specific school catchment will pay a premium that no algorithm would predict. A buyer who grew up in the area might pay more for sentimental reasons.

Agents and appraisers bring market sentiment into their valuations. If we know there’s strong demand from downsizers in a particular suburb, or that young families are moving into an area because of new infrastructure, we adjust our expectations. AI models lag on these qualitative shifts until they show up in actual sales data months later.

I worked with Team400 on a project exploring how AI could better incorporate these softer factors into property analysis. The challenge is that many of the most important variables aren’t quantifiable in ways that machine learning models can process. How do you train an algorithm to recognize “the best street in the suburb” or “the house every buyer wants when they see it”?

Where the Tech Is Heading

The next generation of AVMs is incorporating image analysis from listing photos. In theory, algorithms can assess renovation quality, finishes, and presentation by analyzing pictures. Early results are mixed. The models can identify granite countertops versus laminate, but they struggle with assessing taste, style coherence, or whether a renovation adds value or creates an over-capitalization problem.

Some platforms are trying to incorporate local agent input into their models. An AI valuation might flag a property and then prompt local agents to provide qualitative assessments that adjust the automated figure. It’s a hybrid approach that acknowledges the limitations of pure automation.

Predictive valuations are another area of development. Instead of just telling you what a property is worth today, these tools try to forecast what it’ll be worth in 12 or 24 months based on planned infrastructure, zoning changes, and demographic trends. I’m skeptical of this. Even human experts are terrible at predicting short-term property price movements, and adding AI doesn’t magically make prediction easier.

Practical Implications

If you’re a seller, don’t rely solely on an online valuation tool to set your price. Get a human appraisal from someone who knows the local market. An experienced agent will walk through your property, understand its specific features, and compare it to recent sales with context that an algorithm can’t provide.

For buyers, AI valuations can be a useful starting point for research. If you’re looking at a property priced at $1.9 million and multiple AVMs are suggesting it’s worth $1.6 million, that’s worth investigating. But don’t assume the AI is right and the seller is wrong. There may be features or circumstances that justify the premium.

Investors using these tools for portfolio analysis should understand their limitations. Aggregated data is useful for identifying trends, but individual property decisions still require due diligence that goes beyond what an algorithm can provide.

The Verdict

AI valuations are improving, but they’re not replacing human expertise anytime soon. They’re a tool, not a solution. For standardized properties in data-rich markets, they’re reasonably accurate. For unique properties, complex situations, or markets with rapidly changing dynamics, they fall short.

The best approach is hybrid: use AI tools to process large amounts of data quickly and identify patterns, then apply human judgment to interpret context, assess quality, and make final decisions. That’s how we’re using it at our agency, and it seems to be the direction the industry’s heading.

Anyone claiming AI can fully replace property appraisers is selling something. The technology’s valuable, but it’s complementary to expertise, not a replacement for it.