When AI Starts Choosing the Comparables

An agent preparing for an appraisal opens the comparable-sales report and finds three neat matches already waiting. They sit inside the right radius, sold recently and look close enough on bedrooms, bathrooms and floor area.

The agent also knows that another apartment in the same complex sold quietly last month. The price never appeared in the public feed, although the office handled the transaction and holds the result. Its floor plan, condition and position in the building make it stronger evidence than two of the sales in the report.

Adding that sale may change the price range before the owner meeting begins. The software worked from the transactions it could see, and the strongest comparison sat outside that set.

A property report can look complete because every field is filled and every sale has a photograph. Its first limitation may sit further back, in the evidence that was available to retrieve and the rules used to rank it.

Before the number appears

RICS opened a consultation on the second edition of its comparable-evidence guidance on 8 July. The consultation remains open until 18 September and includes a new section on artificial intelligence and automated valuation models.

The draft says that the output of an AI or AVM system does not, by itself, constitute a valuation. An AVM can contribute evidence, while the valuer decides how much weight to give it. The valuer should understand the process, the data and the limitations, take care when several models rely on similar evidence, and disclose the use of AI or an AVM when arriving at an opinion of value.

Those points belong to professional valuation practice, and the document is still a consultation draft. An Australian estate agent preparing an estimated selling price is doing different work under state rules. The wider direction is still useful because RICS has placed the evidence behind the model alongside the output it produces.

Verification receives as much attention as calculation. The draft says comparable transactions should be checked beyond a headline figure, particularly where they contribute heavily to the opinion of value. It also calls for a record of the research and data sources used to verify the evidence.

Comparable selection is also developing into a model task of its own. A research paper posted to arXiv in March, Retrieval-Enhanced Real Estate Appraisal, tested a system that learns which comparable properties to retrieve. The researchers combined property similarity with geography and trained the selection and estimation components together. Across five datasets in the United States, Brazil and France, their approach used fewer carefully selected comparables and fewer model parameters while remaining close to the performance of the comparison models they chose.

There is no Australian dataset in the study and no basis for assuming that a mainstream Australian CMA product uses the same method. The paper establishes a narrower point: software can learn a selection policy instead of relying only on a fixed radius, date range or feature filter.

The benefit described by the researchers is easier human examination as well as a smaller model. A short, relevant set can be more useful than a large list. The quality of that set still depends on the transactions and property attributes available to the system, and on what its training has taught it to retrieve.

Cotality's Australian products show how closely property search, comparable-market analysis, appraisal preparation and automated estimates now sit in the ordinary agent toolset. Its own AVM disclaimer also sets out the boundary plainly. An estimate is derived from available data without a physical inspection and may not account for condition, improvements, renovations, views, aspect and other features or risks that can affect value.

The Australian Property Institute found a similar boundary in professional attitudes. In an April research note covering more than 300 property professionals, about two-thirds of respondents were open to AI being used in the formulation of valuation reports. The most common view was that it should be confined to certain functions. Respondents were least comfortable with valuation calculations and more comfortable with graphs and document summaries, although the public note provides only limited survey methodology.

That division becomes harder to hold when a supporting task selects the evidence used in the calculation. A system can leave the final price to the person and still influence the range by deciding which transactions arrive at the top of the file.

Victoria puts the choice with the agent

Consumer Affairs Victoria gives this issue a direct agency setting. An agent's estimated selling price must be reasonable and take account of the sale prices of the three properties the agent considers most comparable, unless the agent reasonably believes that fewer than three qualifying sales exist. The estimate sits in the sales authority. CAV explains to sellers that it is neither a professional valuation nor a guaranteed selling price.

For a property in metropolitan Melbourne, a comparable must have sold within the previous six months and sit within two kilometres. Outside metropolitan Melbourne, the limits are 18 months and five kilometres. Those boundaries create the eligible area. They do not make every sale within it equally useful.

The November 2025 Director's Guidelines require agents to consider the differences that change a comparison in practice: construction, age, style and layout, renovation status, floor and land size, bedrooms, bathrooms, parking, special features, zoning, frontage, neighbourhood, school zone, sale date and build status. They also direct attention to substantially similar properties, including another unit in the same complex or a house in the same estate by the same builder.

The missing apartment from the opening therefore cannot be dismissed because it was absent from the software report. CAV says agents must use all the data and information they hold about recently sold properties, including undisclosed sale details, when choosing the three most comparable sales.

The agent still has to decide how the same-complex sale compares with the subject property. Perhaps it has a better floor plan but an inferior aspect. Perhaps its price was affected by an unusual settlement or included a car space the subject does not have. Access to the transaction earns it consideration. It does not settle the weight by itself.

Human selection can be incomplete or distorted too. A person choosing the set still needs evidence, with responsibility for the choice sitting where it can be examined.

Two weeks ago I wrote about retaining an experienced agent's reason for removing a poor comparable. That correction can improve the next appraisal. The current issue begins one step earlier. Before anyone excludes a sale, the software and its data have already defined the candidates the agent is likely to examine.

CAV says agents should keep records of how they selected the three most comparable properties, and it can require information or documents to substantiate the selection. A final Statement of Information shows the chosen sales. The agency's internal record begins with what sat in front of the agent before those three were confirmed.

NSW is moving in the same direction through a different timetable. Some amendments to its property and stock agents laws began on 29 June. A later stage, expected toward the end of 2026 on a date still to be announced, will require agents to consider comparable sold prices under prescribed requirements, retain records when determining or revising estimated selling prices, and prepare a Statement of Information that identifies comparable sales.

Those later provisions are not yet in force. Current NSW guidance already expects an agency to support an estimate and any revision with evidence, including comparable sales, property characteristics, buyer feedback, relevant valuations and other factors affecting price. The coming forms will make more of that price evidence visible to sellers and buyers.

Keeping the selection readable

If software prepares more of the appraisal pack, I would keep a copy of the candidate sales it produced, along with the source and retrieval time. That gives the office a dated view of the evidence the system could see before the agent edited the report.

The approved appraisal file can then hold the final comparable sales and a property-specific note for any material addition or exclusion. Same complex is only a starting point. The note should identify the floor-plan difference, condition, outlook, zoning, frontage, transaction circumstance or other feature that changed the agent's view.

When an undisclosed result is used, the agency also needs to handle the information within the privacy and disclosure rules that apply. The internal file can record that the evidence was considered without turning every piece of confidential information into public copy.

During the campaign, written offers, buyer feedback and newly settled sales can be connected to the estimate that was current at the time. If the price view changes, the file should show the new evidence, the revised estimate and the person who approved it. Six weeks later, another staff member should be able to follow the sequence without reconstructing it from email, memory and an overwritten CMA.

That record is wider than the exact file described by Consumer Affairs Victoria or the coming NSW forms. It is my operating recommendation for an agency using software-prepared evidence: retain the candidate set and source, the selected sales, the material changes made by the agent, and the later evidence that moved the estimate.

It also provides a better way to assess appraisal software. The demonstration usually ends with a polished CMA and a fast price estimate. I would ask to see the point before that: which transactions were searched, what data is missing or delayed, why one sale ranks above another, whether the agent can add a result the office holds, and whether that history remains available after the PDF is produced.

An experienced agent will sometimes disagree with the system. The file should make that disagreement visible, retain the changed set and help the next person inspect it. When only the final three sales survive, no one can reliably tell whether the selection was careful or merely convenient.

The same-complex sale in the opening may lift the estimate, lower it or make no material difference once its circumstances are understood. Its value to the file begins with being seen.

The software may prepare the first set before the agent leaves the office. The agency should still be able to open the record later and show which sales shaped the estimate, which evidence changed the selection and what new information changed the price view.