AI pricing tools have become a normal part of the listing conversation. Vendors expect them. Buyers cite them. Most agencies in Sydney are now running at least one in the background of every appraisal. The interesting question is whether they actually agree with each other, and whether the gap between them is narrowing as the models improve.

Over the last fortnight I ran six AI pricing tools across the same five listings — three in Sydney’s inner west, one on the lower north shore, and one in the eastern suburbs. The pattern that came back is worth talking about.

What I tested

I won’t name the products. The market is small enough that singling them out would generate more noise than signal. The mix included two from the major data providers, two specialist PropTech startups, and two general-purpose AI tools where I fed them comparable sales and asked for an opinion.

For each property I asked the same question: a recommended sale price range and a confidence band. I gave each tool the same comparable sale set, the same property attributes, and the same recent market context.

The results were messier than I expected

On the inner-west houses, the spread between the highest and lowest AI estimate was 14% on one and 19% on another. That’s not a rounding error. On a $2.4 million house, a 19% spread is over $450,000. Vendors are not going to accept “the AI says somewhere between this and that, plus or minus half a million.”

The lower north shore listing was tighter — about 6% spread. The eastern suburbs property, which had unusual features (water views, heritage listing, awkward block), was the worst. Two of the tools returned ranges so wide they were effectively unusable. One of them flagged low confidence, which is at least honest. The other did not.

Where the tools agreed

Interestingly, the tools agreed pretty well on the direction even when they disagreed on the absolute number. If one tool said the property was priced at the upper end of its area’s recent comparables, the others usually agreed. Same with sub-comparable pricing.

This is consistent with what I’ve been seeing in agent practice. The tools are useful as a sanity check on positioning. They are less useful as a source of a final number you’d take to a vendor.

Confidence bands are still oversold

Most of the tools display a confidence band — sometimes a percentage, sometimes a low/medium/high indicator. In my testing the confidence bands were often optimistic. A tool would show “high confidence” on a property where the comparable set was thin or where one outlier sale was clearly distorting the mean.

The honest answer is that AI pricing tools work best in suburbs with high transaction volumes and homogeneous housing stock. They struggle in suburbs where every house is different, where prestige listings dominate, or where the recent sales mix has shifted.

What I do with the outputs

I treat the AI tool output the way I treat a CMA from a junior agent. It’s an input. It surfaces comparables I might not have thought of. It puts a number on the table that I can argue with. But it doesn’t replace walking through the house, reading the strata report, talking to the neighbours, and understanding the buyer pool.

For the more complex prestige listings, I’ve found it useful to bring in AI strategy support when the conversation moves from “should we use these tools” to “how do we integrate AI into our actual workflow without it taking over.” The agencies I’ve seen adopt AI well are the ones where someone has thought hard about which decisions the AI is allowed to influence and which it isn’t.

The advice I give vendors

When a vendor turns up with their own AI estimate from one of the consumer-facing tools, I don’t dismiss it. I show them my own AI tool output as well, and I show them the spread. The honest message is that the AI tools are getting better, but they are still not at the point where you can hand them the appraisal and walk away.

The vendors who push back the hardest on a recommended price are usually the ones citing one specific AI number. The vendors who hear the spread tend to engage more constructively with the actual comparables.

Where I expect this to go

By the end of 2026, I expect at least one of the major data providers will have a pricing tool that produces consistently tighter ranges in standard suburbs. The differentiator will be confidence calibration — being honest about when the model doesn’t know.

The tools that survive will be the ones built by teams that understand both the data and the agent workflow. The tools that struggle will be the ones built by teams that assume the data is enough.

For now, the practical advice is unchanged. Use the AI pricing tools. Don’t trust any single output. Look at the spread across two or three. And when the spread is wide, that’s the model telling you the market is genuinely uncertain — which is usually the most important information you’ll get out of the exercise.