When the algorithm picks the price: APRA's AI letter puts AVMs in the spotlight

On 30 April 2026, the Australian Prudential Regulation Authority (APRA) wrote to every regulated bank, insurer and superannuation trustee in the country with a blunt message: AI adoption has outrun AI governance, and the regulator is preparing to act on it.

The letter — signed by APRA Member Therese McCarthy Hockey and titled APRA Letter to Industry on Artificial Intelligence (AI) — followed a deep-dive engagement with a sample of the country's largest financial institutions in late 2025. APRA found that boards are still developing the technical literacy needed to oversee AI, that assurance practices are not keeping pace with deployment, and that supplier concentration on a handful of AI providers is creating risks the system is not adequately managing. Where entities fail to address these gaps, APRA wrote, it will take "stronger supervisory action and, where appropriate, pursue enforcement."

For Australia's property sector, the letter matters for one specific reason: APRA explicitly named loan application processing as an area where AI is being deployed at scale. That is the workflow where automated valuation models (AVMs) now sit at the centre — and where the regulator's concerns about model risk, supplier concentration and assurance gaps land hardest.

How AVMs quietly became the bank valuation

A decade ago, a residential property valuation in Australia meant a qualified valuer driving to the address, walking the rooms, and signing a report. Today, for the majority of refinances and a growing share of purchase transactions, no human ever sees the property. The bank pulls an automated valuation from one of a small group of providers, decides the model's confidence interval is acceptable, and proceeds.

Three providers dominate. Cotality — the brand under which CoreLogic Australia now trades — markets its AVM as covering more than 96% of Australian residential properties with real-time estimates, and supplies most of the big four banks. PropTrack, owned by REA Group, offers an AVM and a Desktop Valuer product used by lenders and brokers alongside the realestate.com.au portal data feeds. Domain Group publishes its own Home Price Guide estimates and licenses data to a range of financial institutions. Beyond the AVMs themselves, traditional valuation firms such as Opteon and Herron Todd White have integrated digital and AVM-assisted workflows into their full-service offerings.

The economics are obvious. A physical valuation costs the lender several hundred dollars and takes days. An AVM returns a price and a confidence band in seconds, at near-zero marginal cost. For a refinance under a comfortable loan-to-value ratio (LVR), the bank's risk team is usually content to rely on the model. The Australian Banking Association has previously noted that desktop and AVM-based valuations now account for the majority of residential mortgage valuations processed by the major banks.

The takeaway: the AVM is no longer a back-office curiosity. It is the default valuation for most Australian home loans, and it is built on AI models supplied by a small number of vendors.

The letter also flagged that AI is being used in software engineering, claims triage, fraud disruption and customer interaction — but the loan application reference is the one with the most direct read-through to property.

Where AVMs go wrong — and who pays

The friction the consumer feels most often is the "valuation shortfall": a property bought at auction for one price, valued by the bank's AVM at a lower figure, leaving the buyer to make up the gap from cash or renegotiate the loan. Industry data on shortfalls is patchy because lenders rarely publish it, but mortgage brokers consistently report shortfall complaints as one of the most common pain points in current home-loan workflows.

AVMs work well where the underlying comparable-sales data is deep and homogeneous — a 1990s three-bedder in a well-trafficked suburb is the model's comfort zone. They work less well where comparable sales are thin, the property is unusual, or the local market has moved sharply since the most recent recorded sales. In those cases the AVM still produces a number; it produces one with a wider confidence band that the borrower never sees.

The accountability path when a number is wrong is also less clear than with a human valuer. A registered valuer signs a report and can be questioned, and in extreme cases sued. An AVM produces an output through a proprietary model, and the dispute path runs through the bank's review process, not the model provider.

The letter also flagged that AI is being used in software engineering, claims triage, fraud disruption and customer interaction — but the loan application reference is the one with the most direct read-through to property.

The legal and regulatory layers around AVMs

APRA's AI letter does not exist in isolation. AVMs in Australia already sit inside several overlapping frameworks.

Prudential Standard APS 220 Credit Risk Management governs how authorised deposit-taking institutions assess credit risk on residential mortgages, including valuation policies. APRA's APG 223 Residential Mortgage Lending practice guide explicitly defines AVMs and sets expectations for how they should be used in lending decisions. The International Valuation Standards (IVS 2025), adopted by the Australian Property Institute, treat AVM outputs as one of several valuation methodologies with disclosure requirements around model assumptions and confidence levels. And the federal government's Privacy Act reforms — which introduced tighter rules around automated decision-making — are now layering consumer protection obligations on top of the prudential framework.

The Australian Securities and Investments Commission (ASIC) has separately flagged AI governance gaps in financial services as a supervisory priority, including in credit and advice. The Council of Financial Regulators, of which APRA, ASIC and the RBA are members, has been coordinating policy on AI in financial services across the sector.

For lenders, the practical effect is that an AVM is no longer just a procurement question or a credit-risk question. It is a model-risk question, a privacy question, a supplier-concentration question and, increasingly, a board-governance question.

What it means for property buyers and investors

For homebuyers and existing mortgage holders, the practical takeaways from APRA's letter are indirect but real. Banks will be under more pressure to demonstrate they understand the AVMs they rely on, which is likely to mean more conservative use of AVMs for non-standard properties and a renewed willingness to commission full physical valuations where the model's confidence is low. Borrowers buying unusual properties — warehouse conversions, regional acreage, high-end apartments — may find that lenders defer to a physical valuation more often than in 2024.

Borrowers facing an AVM shortfall now have a stronger basis for asking the lender to commission a second valuation, and to disclose the basis on which the AVM was relied upon in the first place. The Privacy Act reforms also give consumers stronger rights to information about automated decision-making, although the application of these rights to AVM-driven credit decisions is still being tested in practice.

For property investors with multi-loan portfolios, the supplier-concentration concern is the one to watch. If APRA pushes banks to diversify their AVM dependencies, the consistency of valuations across a portfolio — already an issue when different lenders use different providers — may become more variable, not less.

The takeaway: AI is not going to be unwound from the home-loan process. But APRA has made clear that the era of treating AVMs as a low-effort cost-saver is ending.

Looking ahead

APRA's letter described its current AI work as "guidance based on current observations" rather than new prudential standards. The regulator said it is "currently finalising its forward plan with regards to supervision of AI risks" and explicitly left the door open to further policy action.

The most likely near-term moves include updated prudential guidance specifically addressing AI in credit decisioning, expansion of CPS 230 Operational Risk Management — itself finalised on the same day — to cover AI supplier concentration, and joint policy work with ASIC on automated decision-making in consumer credit. The federal government's broader digital and AI policy agenda, including the Corporations Amendment (Digital Assets Framework) Bill passed in April 2026, has also raised the regulatory tempo across financial services.

For now, AVMs will continue to price most Australian homes. But the institutions buying that pricing have been put on formal notice that the regulator will be looking far more closely at how those numbers are generated, monitored and acted on — and that the days of treating the algorithm as a black box are over.

Key takeaways

• APRA's 30 April 2026 letter is the first published AI-specific guidance from the prudential regulator, and warns of enforcement where regulated entities fail to manage AI risks adequately.
• The four governance failings APRA identified — security, governance maturity, supplier concentration, and assurance — all apply directly to AVMs used by lenders.
• Cotality (CoreLogic), PropTrack (REA) and Domain dominate the Australian AVM market, with Cotality alone covering more than 96% of residential properties.
• Borrowers facing valuation shortfalls have stronger grounds, post-letter and post-Privacy Act reform, to seek a second valuation and disclosure about the basis of the AVM.
• More targeted AI-specific prudential guidance is expected from APRA in the coming months.