I’ve been tracking automated valuation model (AVM) accuracy across Sydney for the past eighteen months, and the improvements are real. The gap between AI-generated estimates and actual sale prices has narrowed considerably since I first started writing about this topic.

But “better” doesn’t mean “ready to replace appraisals.” And understanding exactly where AVMs succeed and where they fall short is becoming a critical skill for agents who want to maintain credibility in listing presentations.

The Numbers Tell an Interesting Story

CoreLogic’s AVM accuracy for Sydney metro properties has improved from roughly ±8% median error in early 2025 to around ±5.5% as of February 2026. PropTrack’s model shows similar gains. For a $1.5 million property, that’s still a potential swing of $165,000 — which is a massive range when you’re trying to set vendor expectations.

The improvements aren’t uniform, either. AVMs perform best for freestanding houses in established suburbs with high transaction volumes. Think Baulkham Hills, Castle Hill, Cronulla — places where there are dozens of comparable sales each quarter and the housing stock is relatively homogeneous.

They’re weakest for unique properties, strata complexes with mixed configurations, and anything in a thin market. A three-bedroom apartment in a boutique block of eight in Mosman might have no genuinely comparable recent sale. The algorithm extrapolates, and the results can be wildly off.

What’s Driving the Improvement

Three things have changed since I last wrote about this.

Better data inputs. AVMs now incorporate more granular data — not just bedrooms, bathrooms, and land size, but renovation quality indicators, aspect analysis from satellite imagery, and even proximity scoring to amenities that actually affect prices (good schools, transport, specific retail precincts).

Larger training datasets. The post-COVID transaction volume gave these models millions of additional data points across varied market conditions. They’ve now “seen” a rate hiking cycle, a correction, and a recovery. That makes their predictions more robust across different market phases.

Feedback loops. Several AVM providers now incorporate agent feedback and actual settlement data to continuously refine their models. Team400 has been working with property data firms on exactly this kind of machine learning pipeline — using real-world outcomes to train better predictions.

Where Agents Still Win

I ran a small experiment last quarter. I asked three experienced agents to appraise fifteen properties across different Sydney suburbs, then compared their estimates to the AVM predictions and the eventual sale prices.

The agents’ median error was ±3.2%. The best-performing AVM came in at ±5.1%.

The agents were consistently better at:

• Strata properties where building-specific factors (special levies, management quality, common area condition) dramatically affect pricing
• Properties with recent renovations where the quality and taste of the renovation matters as much as the fact it was done
• Transitional suburbs where buyer demographics are shifting and historical sales data doesn’t capture the current demand profile
• Days on market prediction — agents were significantly better at estimating not just price but how long it would take to sell

The AVMs performed better than agents on exactly one category: deceased estates and executor sales where the property had been unmaintained. Agents tended to over-discount for condition issues, while the AVMs more accurately priced the underlying land and location value.

Practical Implications for Listing Presentations

Here’s what I tell the agencies I consult with.

Use AVMs as a starting point, not a conclusion. Pull the CoreLogic, PropTrack, and Domain estimates before every listing appointment. They give you a data-anchored range that’s useful for framing the conversation.

Show vendors the limitations transparently. Don’t pretend AVMs don’t exist — vendors already check them on Domain and realestate.com.au. Acknowledge the estimates, then explain specifically why your appraisal differs. “The algorithm doesn’t account for your north-facing aspect and the kitchen renovation” is a concrete, defensible statement.

Track your own accuracy. Every quarter, compare your appraisals against actual settlement prices. If you’re consistently within 3%, you’ve got a powerful story to tell vendors about why your judgment adds value beyond what a website can offer.

Don’t fight the technology. The agents who’ll thrive are the ones who use AVMs to inform their thinking while adding the human layer that algorithms genuinely can’t replicate. The days on market insight alone — knowing whether a property will sell in two weeks or sit for six — is worth more to most vendors than the price estimate itself.

Where This Goes Next

Within two years, I expect AVM accuracy for standard properties in high-volume suburbs to match experienced agents. That’s not a threat — it’s an opportunity to redirect your value proposition toward the complex, high-stakes transactions where human expertise genuinely matters.

The clearance rates in Sydney have been hovering around 65-68% through early 2026. In that kind of balanced market, accurate pricing is everything. Getting it wrong by even 5% can mean the difference between selling under the hammer and a property that passes in and lingers.

The smart play is to get comfortable with these tools now, while you still have a clear edge. Because the gap is closing, and the agents who understand both the technology and its limitations will be the ones vendors trust most.