I’ve been watching automated property valuations improve for years now, and they’re genuinely impressive in some respects. But I’m still seeing situations where the AI spits out a number that’s wildly off base, and everyone involved acts surprised.

The gap isn’t usually the algorithm’s fault. It’s about what gets fed into it.

What AI Gets Right

The models are excellent at macro trends. They’ll pick up suburb-level price movements, identify which areas are heating up, and spot patterns across thousands of transactions. If you’re looking at median prices or broad market shifts, the data’s solid.

They’re also getting better at factoring in obvious features. Land size, bedrooms, bathrooms, parking spots, proximity to transport. All of this gets weighted reasonably well, and for a standard property in a well-traded suburb, you’ll often get a valuation that’s within 5-10% of reality.

For some use cases, that’s perfectly fine. Lenders doing portfolio risk assessments don’t need precision down to the last $10,000. But if you’re a vendor trying to decide on a listing price? That margin matters.

Where It Falls Apart

The problem is nuance. And Sydney properties are full of it.

I had a client last month who got an automated valuation on a terrace in Newtown. The AI pegged it at $1.65 million. Reasonable, based on recent sales of similar-sized terraces nearby. Except this one had an illegal rear extension with no council approval, dodgy waterproofing, and structural issues that would cost $200K minimum to fix properly.

The algorithm didn’t know that. It saw square meterage and said “sounds good.”

On the flip side, I’ve seen properties undervalued because they have features the model doesn’t recognize. A north-facing garden in an inner-city terrace, high ceilings in a renovation, quality fixtures that don’t show up in land registry data. These things add real value, but they’re invisible to an algorithm working off public records.

Street-level differences matter too. Two streets in the same suburb can have completely different buyer appeal, but if the data’s thin, the AI treats them as equivalent. I’ve seen this particularly in areas undergoing gentrification, where one end of a suburb is booming and the other’s still lagging.

The Trust Problem

Here’s the tricky part: vendors are starting to treat these automated numbers as gospel. They’ll get a CoreLogic estimate or a Domain AVM and anchor to that figure, sometimes against their agent’s advice.

I get it. The number comes from “data,” so it feels objective. But objectivity and accuracy aren’t the same thing. If the data’s incomplete or the model can’t account for property-specific factors, you’re getting a precise answer to the wrong question.

This is where agents still add value. A good agent knows the micro-market, has seen inside comparable properties, understands what buyers in that area actually care about. That’s not something you can automate away, at least not yet.

What Actually Works

I’m not anti-AI here. I use automated valuations all the time, but as a starting point, not an endpoint. They’re useful for:

• Getting a rough range quickly. If a vendor thinks their house is worth $3 million and the AI says $1.8 million, that’s a reality check worth having.
• Tracking market movements. Month-to-month changes in AVMs can show you which suburbs are trending up or down before it shows up in settled sales data.
• Spotting outliers. If a property’s valuation is significantly different from its neighbors, that’s worth investigating. Sometimes it’s an error; sometimes it’s a genuine insight.

But you still need a human who knows the property to interpret the number. That means either an experienced agent or an independent valuer, depending on the stakes.

Some firms are getting smarter about this. Team400, for instance, has been working with agencies to build tools that combine automated data with local agent knowledge, which feels like the right direction. You want the speed and consistency of AI, but with the context that prevents obvious mistakes.

The best approach I’ve seen is a hybrid: use the automated valuation to set boundaries, then apply local knowledge and property-specific factors to refine it. That gets you closer to a number you can actually defend.

What’s Coming

The models will keep improving. More granular data, better image recognition (so the AI can “see” renovations and condition from photos), integration with council records to flag issues like unapproved extensions. All of that’s either here already or coming soon.

But I don’t think we’re heading toward a world where you just plug in an address and get a perfect answer. Real estate’s too messy for that. What’s more likely is that the AI handles the bulk of the analysis, and humans focus on the edge cases and judgment calls.

For now, if you’re a vendor, treat automated valuations as one input among many. Get a few different estimates, talk to agents who know your area, and if the stakes are high enough, pay for a formal valuation from a certified valuer. The $600 or so it costs could save you from listing too high and sitting on market, or too low and leaving money on the table.

And if you’re an agent, don’t be defensive about this stuff. AVMs aren’t going to replace you, but they are going to change what you’re hired to do. The agents who embrace the data and learn to work with it will have an edge over those who dismiss it outright.

The tech’s a tool. Whether it helps or hinders depends on how you use it.