AI in property valuation: the evidence, the design rules, and what it could become

The best Australian evidence on computer vision in property valuation was funded by the valuation profession itself, and it is almost never cited in the conversations valuers are currently having with vendors.

In 2024 the Australian Property Research and Education Fund, the API’s research arm, funded a Queensland University of Technology team (Associate Professor Viet-Ngu Hoang, Dr Kien Thanh Nguyen and Dr Andrea Blake) to test a specific question: can a computer looking at pictures of a property improve a statistical estimate of what that property is worth?

The result only means something against what it improved on. That baseline is the hedonic pricing model behind most automated valuation models, predicting sale price from the attributes that fit in a spreadsheet plus the usual neighbourhood factors. It has never seen the property. Two houses with identical attributes are identical to it, whether one has been renovated to a high standard and the other is tired and fronting a main road.

The team trained convolutional neural networks on aerial and street-view imagery to generate visual desirability indicators, added those to the model as extra variables, and tested across 128 Brisbane suburbs on three single-year datasets from 2018 to 2020. Street-view imagery cut the model’s typical prediction error, measured as root mean square error, by 11 to 21 per cent. Aerial imagery cut it by 20 to 29 per cent.

On a model typically out by $100,000, aerial imagery brings the miss down to somewhere between $71,000 and $80,000. The dollar baseline is illustrative, since the study reports the proportional improvement rather than a figure. The model does not become accurate. It becomes meaningfully less wrong, and the improvement came from pictures alone.

The finding that should interest a valuer is which pictures won. Aerial beat street-view, comfortably.

The street-view image is the front elevation, the photograph a person would instinctively reach for, and it carried less signal than the view from directly overhead. Nobody attends a property to take that photograph. It already exists, for every property in the state, in imagery a practice can licence today.

These are numbers from an academic team with nothing to sell. They are the strongest Australian evidence in this area, and they deserve to be where a practice starts.

The evidence measures a different thing than the demonstration

The QUT work is about mass valuation. It improves a statistical model that estimates value across a large number of properties. The imagery that carried the most signal was aerial, which is to say imagery of the roof, the block, the setbacks, the tree canopy and the neighbours, captured from above.

What is typically demonstrated to a valuation practice is something quite different: a tool that reads the photographs an inspector took at a property and produces a description, a schedule, or a draft report. That is individual report production, not mass valuation.

The distinction is not academic. Nothing in the QUT result tells you whether a model reading twenty inspection photographs produces a schedule a valuer can rely on for one property. Those are different tasks with different failure modes, and we have not found a published independent Australian benchmark for the second.

A valuer sitting through a demonstration is entitled to ask which of the two the vendor has evidence for. In most cases the answer will be neither, and the honest version of that answer is a case study rather than a benchmark.

What our own production work shows

PAC has built and run vision systems in two adjacent domains: an automated listing copy engine that reads property photographs, and research into AI-assisted inspection reporting. Both are our own work, and the findings below are ours rather than any client’s.

Six of them transfer to a valuation report, and they build on each other.

Detection is strongest where liability is weakest. Systems reading imagery perform well on visually distinct features and conditions, and degrade on anything subtle, internal, or occluded. In inspection work that inverts the risk profile, because liability concentrates on the defects that were present, within scope, and missed, not the obvious ones. The same shape applies to a valuer: a model will reliably tell you a kitchen has been renovated and will not reliably tell you the quality of what sits behind the cabinetry. The tasks a vision system does best are the tasks a competent professional was never going to get wrong. We covered this pattern across three professions in AI in regulated professional work, Mid-2026.

A defined data model is what removes most of the room to hallucinate. Vision models capture non-deterministically. Ask one to describe a kitchen and it produces prose, and prose is where invention lives, because nothing in the task tells the model what it is not permitted to say.

DreamCopy does not work that way. Behind the copy sits a property data model: a fixed schema of fields, each with a defined type. The model is not writing about the property. It is populating a record, and the copy is generated afterwards from the populated record rather than directly from the photographs. A field the model cannot determine returns a flag rather than a guess.

For a valuation practice the data model already exists. It is the schedule. The firm’s report template, with its fields and its house taxonomy for construction, condition and finish, is precisely the schema a vision system should be populating.

The schema decides the photography, and the photography decides the cost. Once the fields are defined they dictate what the system has to be able to see.

DreamCopy processes photographs at low resolution. It can, because the task only needs a general feel for the saleable qualities of a room: light, space, flow, the broad standard of the fit-out. Low resolution keeps the upload fast at the point of capture, keeps per-image processing cheap, and lets an entire property run through for a trivial sum.

A valuation schedule asks for something else. Separating natural stone from engineered stone, judging the quality of a tile, reading an appliance brand from the badge on a door: each needs more photographs and higher-resolution ones, framed deliberately on the thing being recorded rather than on the room it sits in.

Three things move at once as a result, and none of them are software problems. Upload time at the property rises, on whatever connection happens to be available there. Storage and per-image processing costs rise, and not marginally. And past a certain fidelity requirement the architecture itself changes shape. A single pass across a set of room photographs stops being the right design, and the pattern becomes staged: a cheap low-resolution pass identifies what matters, then a targeted high-resolution pass reads it.

What a practice is really designing here is the capture routine: what gets photographed, how closely, and to what standard, on every job.

A model can narrow a judgement without making it, and that is where the time is. Consider the attributes that actually move a valuation figure at the finish level: stone benchtops, premium tiles, solid hardwood flooring, premium appliances. A photograph cannot settle any of them. Natural stone and engineered stone are indistinguishable in a listing image. So are solid hardwood, engineered timber and a good laminate.

Those attributes still have to be hand-noted, and no architecture changes that.

What a model can do is turn origination into selection. Rather than leaving a blank field, it proposes the plausible candidates for that benchtop with what it can see behind each one, and the valuer picks the right one or overrides it entirely. Reading four options and choosing is faster than recalling the taxonomy and typing, and it is faster across every field on every job. The speed comes from the valuer deciding more quickly, not from the model deciding instead.

Supplied facts have to beat inferred facts, structurally, not by convention. The rule bites hardest where the firm already holds the answer. In one production run our system correctly hedged that a property sat close to a suburb boundary. It had found real ambiguity and said so. The next stage, with no authoritative address to check against, flattened that hedge into one confident answer and chose the wrong suburb.

That failure is worth sitting with, because the model was not wrong. The model was right, and calibrated, and the architecture destroyed the calibration on the way to the output. No better model would have prevented it. What prevented it was a rule: the supplied address wins, always, over anything inferred from imagery.

For a valuation practice that covers everything the firm already holds authoritatively. Title particulars, land area, zoning, planning overlays, sales evidence. Those are supplied facts, and an inference must never overwrite one.

None of this is one model call. It is an additive pipeline, and every stage needs its own step and its own context. The most common misconception about a system like this is that a vision model reads the photographs and a report comes out the other side. What actually runs is a sequence, and each stage exists as a separate step because merging any two of them measurably degrades the result.

In DreamCopy the sequence is a vision stage that reads the photographs; an extraction stage that takes the agent’s own notes and pulls out what only the agent knows; a location stage that resolves the address and brings in locality and market data; the data model, where all three land as one structured record; and then the reporting layers that generate copy from that record.

Each stage carries its own context and its own quality bar, because a model asked to perceive and compose in the same breath does neither well, which is the finding behind how we turned generic AI into a specialist. Analysis has to finish before generation starts, and every stage after the first works from a record rather than from raw material.

The valuation analogue maps almost directly. Vision reads the property. An extraction stage takes the valuer’s field notes or dictation, which is where everything a photograph cannot carry enters the system. A data stage brings in title, zoning, planning overlays, comparables and sales evidence, which is the mature part of the valuation stack and the part the practice already owns outright. The schedule is the data model. The report is the final layer, generated from a confirmed record.

The design rules that follow

Those findings collapse into a short set of rules.

1. Define the data model first. The schedule is the schema. Nothing else can be designed until the fields, types and permitted values are settled.
2. Let the schema set the photography. The fields decide what has to be photographed and at what fidelity. Settle that early, because it sets the field routine on every job.
3. One stage per job, each with its own context. Perception, extraction, data, record, report. Analysis finishes before generation begins.
4. Authoritative facts win structurally. Anything the practice already holds overrides anything inferred from an image, enforced in the architecture.
5. Present alternatives, never resolve them. Where no authoritative fact exists, the ambiguity goes in front of the professional as a choice. A low-confidence output must never be flattened into a confident one downstream.
6. Aim the detection target low. Put the right shortlist in front of the professional. Do not attempt the diagnosis.
7. The professional stays the author, visibly. Not as a concession to regulation, but because it is what makes the output signable.

A practice can take that list into any vendor conversation and learn more from the answers than from the demonstration.

What this could become, and what is still unknown

Valuation is unusual among professions in that the data is already mature. A practice holds sales evidence, comparables, attribute data and, in an established firm, two decades of completed reports and expert witness statements. That corpus is the asset, and no vendor has access to it.

The QUT result shows that imagery adds measurable signal to a statistical estimate built on that kind of data. Our pipeline work shows how to get structured, confirmed property records out of imagery, field notes and location data at volume. Put those together and the proposition is legible: a practice that runs the pipeline over its own jobs accumulates confirmed, structured property records against known outcomes, and those records are exactly the input a valuation model of its own would need.

That is a proof of concept proposition, not a product, and the honest list of what is unknown is longer than the list of what is established.

Whether one practice generates enough confirmed records to move a model at all is untested: QUT worked across 128 suburbs and three years of market data, which is a different order of magnitude from a single firm’s annual volume. Nor is it known whether the Brisbane result holds in another market or another property type.

This also sits in territory the standards bodies are actively writing rules for, which we will cover on its own.

Where the benefit actually sits

The useful way to think about all of this is not that a machine does part of the valuer’s job. It is that the effort moves.

The grind at the front of a valuation is real: populating a schedule field by field, transcribing what was seen, assembling the narrative sections, chasing the consistency that a panel or a lender’s quality team will test for. That work is compressible, and the six findings above are an account of how to compress it without putting anything unverified into a signed document.

What is left is the executive end. Weighing the evidence, deciding which comparables genuinely compare, forming the opinion, and defending it. That is where a valuer’s value has always been, and it is where the time goes instead. A practice that runs this well does not produce reports with less judgement in them. It produces reports with the same judgement, applied to more of the job, with less of the day spent on data entry to get there.

What this note does not establish

We have not found a published independent Australian benchmark for vision AI in individual valuation report production. The QUT figures are for mass valuation and should not be read across to report production. PAC’s production findings are our own operational experience, not independently evaluated. The proprietary valuation model described above is a concept with its unknowns named, not a capability anyone has demonstrated.

Where a vendor tells you otherwise, the useful question is which of those gaps they have closed, and with what evidence.

Perth AI Consulting works with professional practices on what AI can and cannot be relied on for, and what has to be documented either way. Start with a conversation.