Our AI Found a $100K Undervalued Property in One Hour. It Took Me Three Days to Believe It.

AI is going to change property investing the same way it changed chess. Not by replacing humans, but by making the humans who use it impossibly hard to compete against.

We have been training our own property-finding models since early 2021, feeding them data from hundreds of our transactions — every purchase price, every valuation outcome, every renovation cost, every rental result. The goal was simple: build a digital version of our team's collective pattern recognition. A tool that does not sleep, does not get distracted, and does not have a bad day 1.

Last month, it found something that made me sit down and stare at my screen for a long time.

The impossible brief

We had a client with a $1.2 million budget and six requirements that, individually, are reasonable. Combined, they sound like a wish list from someone who has never bought property.

One: annual capital growth of 10 percent or better. Two: a commercial-zoned frontage where the client could operate their GP medical practice. Three: residential space at the rear where the client could actually live. Four: if the business failed, the property had to convert to a pure investment generating at least 5 percent rental yield. Five: it had to be a genuine bargain — meaningfully below market value. Six: subdivision or rear development potential.

I looked at that brief and thought: this property does not exist. Six criteria pulling in different directions. Commercial and residential on the same title. High growth and high yield. Development potential and livability. In my experience, you get two or three of those features in a single property. Not six 2.

But we set up the parameters in the tool anyway. One hour later, it flagged a property.

What the AI found (and why I did not believe it at first)

The property had been purchased in 2017 for $1,020,000. It was now listed at $800,000-$880,000. Even accounting for agent underquoting, the realistic purchase price sat around $1,000,000 — essentially the 2017 price, eight years later.

That alone was unusual. Melbourne property does not typically trade sideways for eight years. Something had spooked previous buyers. The listing had been sitting for longer than average.

But the AI had gone deeper than the listing. It had cross-referenced the zoning overlay, the neighbouring properties' development history, and the building's floor plan against its training data from our previous transactions.

Here is what it identified:

The neighbours on both sides had already completed subdivisions — one had developed two dwellings, another had built three. This confirmed that the zoning permitted development and that council had approved similar applications on the same street 3.

The front section of the property was a commercial building suitable for conversion to a medical practice. The rear had a separate residential structure — already built — that could serve as the client's home.

The driveway was 3 metres wide — sufficient for independent rear access, which is a non-negotiable requirement for subdivision or secondary dwelling approval in most Melbourne councils.

And here is where one of the AI models did something I would not have thought to check: it found that the property at number 104 on the same street had been converted into a veterinary clinic under the same zoning. It verified that the zoning code was identical. If a vet clinic was approved, a GP practice would face no additional zoning hurdles 4.

Three AI models, three perspectives

We run multiple assessment models on each flagged property. Each model has been trained with different aspects of our historical data and looks at the opportunity from a different angle.

Model one — the generalist — gave a straightforward assessment: the property met the self-occupation criteria and had rear subdivision potential. Solid but not remarkable.

Model two — trained specifically on commercial property conversions — flagged the commercial potential and noted the 3-metre driveway width as a positive indicator for future rear access. It also highlighted that the front building's layout was suitable for medical or consulting use without major structural modification.

Model three — the deep researcher — went beyond the property itself. It crawled council records and found the number 104 veterinary conversion. It identified the zoning match. It pulled the comparable development applications from neighbours. This is the kind of research that would take a human half a day of council website navigation and phone calls 5.

Combined, the three models painted a picture that was genuinely compelling. This was not a compromise property that sort-of met the brief. It was a near-perfect match across all six criteria.

There was one flag: a Significant Landscape Overlay affecting the rear yard. If any tree exceeded 5 metres in height or had a trunk circumference exceeding 0.5 metres at 1 metre above ground, removing it would require a permit. That could complicate development. The AI flagged it. A human might have missed it in the Section 32 6.

What AI cannot do (and why that matters too)

I want to be clear about something: this tool does not replace a buyer's agent. It replaces the worst part of a buyer's agent's job — the tedious, repetitive scanning of new listings, the cross-referencing of zoning data, the late-night checks for properties that have fallen through and relisted.

The AI does not negotiate. It does not read the body language of an agent at an open inspection. It does not know that a vendor who has just gone through a divorce will accept 15 percent below asking if you can settle in thirty days. Those are human skills that require emotional intelligence and situational judgment.

What the AI does is ensure that no opportunity slips through the cracks. A property can fail a conditional sale at 10pm and be relisted by midnight with a reduced price guide. The AI catches it immediately. By 9am the next morning, our team has already reviewed it and is on the phone to the agent 7.

In a market where good deals disappear within 48 hours, that speed advantage is worth tens of thousands of dollars per transaction.

The other thing the AI does brilliantly is eliminate confirmation bias. Human buyer's agents — myself included — have suburbs they know well and suburbs they overlook. We have property types we are comfortable with and types we unconsciously filter out. The AI has no preferences. It evaluates every listing against the criteria and flags whatever matches, regardless of whether it is in a suburb I have ever heard of 8.

The bigger picture: data-driven property in Australia

We started building this tool because of a simple observation: the Australian property market generates enormous amounts of public data — listings, sales records, zoning maps, council applications, valuation reports — but almost nobody synthesises it systematically.

Most investors make decisions based on Domain searches, word of mouth, and gut feeling. Most buyer's agents add experience and network access on top of that, which is valuable but still leaves massive amounts of data unexamined.

Our approach is different. We have been cataloguing every transaction we complete — purchase price, renovation cost, valuation outcome, rental achieved, tenant quality, maintenance costs — and feeding that data back into models that can identify patterns invisible to any individual human 9.

After 350-plus transactions, the dataset is substantial. The models have learned what a good deal looks like across multiple dimensions: location, land size, zoning, building condition, price relative to comparable sales, rental potential relative to purchase price.

This is not magic. It is the same approach that quantitative traders have used in financial markets for decades. The property industry is just twenty years behind.

For our clients, the practical impact is simple: we find more opportunities, faster, with fewer misses. The AI handles the scanning. Our team handles the judgment. Together, we are better than either one alone.

The client with the six impossible criteria? We found two additional properties that matched even better than the first. One of those is now under contract. The AI sourced the leads. I negotiated the price. And the client is getting a property that meets every single requirement on a brief that I initially thought was unrealistic 10.