What problem does it solve?
Every mortgage, remortgage and property tax reassessment needs a value, and a full, on site appraisal takes a professional's time to inspect the property, research comparable sales and write a report for each individual property. The Royal Institution of Chartered Surveyors, a global professional body for the sector, notes that residential valuation data is often publicly available while commercial property data is often less widely available, which makes automated models less reliable outside standard, homogeneous housing.
Automated valuation models are not new, and not all of them use machine learning; RICS points out that some still apply fixed, rule based formulas. What has changed is how many of them now learn from large, continuously updated datasets of sales, tax records and property characteristics, and how far organizations are willing to let a model's output stand in for a person's inspection.
How does it work?
- Collect the comparables. The model pulls comparable sales, tax assessment and land registry records, property characteristics and location data covering the area.
- Predict and compare. A regression or machine learning model estimates the property's value and identifies the closest comparable properties that have sold.
- Score the reliability. The system scores how confident the estimate is, typically from how closely it tracks the comparable sales it used.
- Route by confidence. High confidence, low risk cases are auto accepted, for example as an appraisal waiver or a direct enrollment at the sale price; low confidence or high value cases are queued for a person.
- A professional reviews the queue. Valuers or appraisers spend their time on the batches the model flagged as uncertain or close to a decision boundary, not on every case.
- Audience
- Employee facing
- Autonomy
- Supervised agent
- Adoption
- Mainstream
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Customer savings | Not pooled | at least USD 2.5 billion | 1 | 1 organization |
Value drivers: Lower cost to serve, Speed and cycle time, Employee productivity, Customer experience.
Indicative value
A residential mortgage lender originating 50,000 loans a year
USD 1.6 million to USD 35 million
Appraisal fee cost avoided per year
How this is calculated
Formula: loans * avmEligibleShare * appraisalFeeAvoided. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Loans originated per year loans, loans per year | 20,000 | 100,000 | Editorial assumption, replace with your own origination volume. |
| Share of loans eligible for an automated valuation instead of a full appraisal avmEligibleShare, fraction of loans | 0.2 | 0.5 | Editorial assumption, replace with your own eligibility policy and loan mix. |
| Appraisal fee avoided per automated valuation appraisalFeeAvoided, USD per loan | 400 | 700 | Editorial assumption for a typical US conventional appraisal fee, replace with your own. |
What it leaves out: Only the appraisal fee the borrower avoids paying, for the lender's loans that qualify for an automated valuation instead of a full appraisal. It leaves out the cost of building and validating the model, the appraisals still needed for higher risk or higher value loans, and any difference in collateral risk between an automated and a fully manual valuation, which none of the deployments on this page report as a single figure.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Valuation Office Agency
United Kingdom · Government and public sector · 2025
The UK's Valuation Office Agency built its own Automated Valuation Model, without an external supplier, to give a first pass value, the closest comparable sales, a reliability score and a Council Tax band for the vast majority of Wales's 1.5 million domestic properties ahead of the 2028 Council Tax revaluation. Valuers focus their time on the batches of properties the model flags as least reliable or closest to a band boundary, and the International Association of Assessing Officers reviewed the model's development process and reported confidence in its quality.
No outcome disclosed.
Fannie Mae
United States · Banking · 2024
Fannie Mae's Desktop Underwriter can issue Value Acceptance, an offer to skip a traditional appraisal, using what Fannie Mae calls a robust data and modeling framework to confirm the validity of a property's value and sale price. Fannie Mae announced that, beginning in the first quarter of 2025, the eligible loan to value ratio for Value Acceptance on purchase loans for primary residences and second homes would increase from 80% to 90%, and it estimates that appraisal alternatives such as Value Acceptance and Value Acceptance + Property Data have saved mortgage borrowers more than $2.5 billion since early 2020.
- Customer savings: at least USD 2.5 billion, since early 2020
"Since early 2020, Fannie Mae estimates the use of appraisal alternatives such as Value Acceptance and Value Acceptance + Property Data on loans Fannie Mae has acquired saved mortgage borrowers more than $2.5 billion."
Claimed by: organization
Riverside County Assessor-County Clerk-Recorder
United States · Government and public sector · 2024
Riverside County's Assessor-County Clerk-Recorder deployed the C3 AI Residential Property Appraisal application for around 460,000 single family homes and condominiums, replacing manual linear regression models used to automatically enroll eligible change of ownership transfers at their sale price. C3 AI describes this as Riverside's initial production deployment, delivered in under six months, meant to demonstrate the application's ability to improve staff efficiency and reduce the complexity of its modeling approach.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Comparable sales, tax assessment and property characteristic data covering the geography
- A documented method to test accuracy and fairness across property types, values and areas
- A defined confidence threshold that decides which properties get an automated value and which go to a person
Systems to integrate
- Automated underwriting system or tax assessment system that consumes the estimate
- Geospatial, land registry or multiple listing service data feeds
- A case queue for the properties the model routes to manual valuation
Complexity: High
Building or buying a model a regulator, lender or model risk team will accept needs deep, clean sales and property data and a documented method for testing accuracy and fairness across property types and areas; the Valuation Office Agency aligned its in house testing to International Association of Assessing Officers AVM standards and had the International Association of Assessing Officers review its model's development process before relying on it.
- 1
Decide where an automated value can stand in for a person
Set the loan to value, price band or property type limits within which an automated value is acceptable, matching what a regulator or an internal model risk function expects.
- 2
Build the confidence score, not just the estimate
Score every estimate's reliability from its distance to comparable sales, so low confidence cases route to a person automatically rather than being accepted at face value.
- 3
Test for accuracy and fairness before launch
Run a ratio study across property types, price bands and geographies, not just an overall error rate, against a recognized mass appraisal standard.
- 4
Keep a professional in the loop for the exceptions
Route low confidence and high value properties to a valuer or appraiser, and feed that person's corrections back into ongoing monitoring of the model.
- 5
Publish the method
A public register entry or a documented internal policy explaining what the model does and why builds the trust an automated valuation program needs from regulators and customers.
Guardrails
- A confidence score below a set threshold always routes to a human valuer, never an automated value alone
- Independent testing against a recognized mass appraisal standard before launch and after every material model change
- Loan to value, price band or property type limits on when an automated value replaces a full appraisal
KPIs to instrument
- Automated valuations issued versus properties routed to a person, split by price band and area
- Accuracy and dispersion of automated valuations against confirmed sale prices or completed manual valuations
- Appeals or challenges to automated valuations as a share of all automated valuations issued
Human in the loop
A qualified valuer or appraiser reviews every property the model flags as low confidence or above a value threshold, and a central analytics or model risk team monitors overall accuracy and fairness against completed sales and manual valuations.
Common failure modes
- Confident but wrong in a fast moving market
- A model trained on past sales lags a market that is moving quickly and prices systematically low or high; monitor the gap between automated valuations and completed sales every month, not only at launch.
- Uneven accuracy across areas and property types
- Data rich urban areas get a more accurate value than rural or unusual properties, so the model quietly serves some customers worse than others; a ratio study by property type and area, not an overall figure alone, is what catches this.
What are the risks and rules?
EU AI Act
Depends on design
An automated valuation model values the collateral, not the person, so it is not itself listed in Annex III; the EU Mortgage Credit Directive treats property valuation (Article 19) and the creditworthiness assessment of the borrower (Article 18) as separate steps, and Article 18(3) says the creditworthiness assessment must not be based predominantly on the value of the property exceeding the amount of credit, or on an assumption that the property's value will increase. The valuation becomes relevant to Annex III point 5(b), creditworthiness assessment of natural persons, only where its output is built into a separate system that evaluates the borrower's creditworthiness, and whether that happens depends on how the lender designs the credit decision, not on the valuation model itself.
Rules that apply
Guidance
- Responsible use of AI case study: valuation (Royal Institution of Chartered Surveyors, Europe). Explains that automated valuation models should support, not replace, a qualified valuer's judgment for high risk valuations such as those informing lending, legal disputes or investment decisions.
- Quality Control Standards for Automated Valuation Models (CFPB, OCC, Federal Reserve, FDIC, NCUA and FHFA, North America). The 2024 US interagency final rule requiring mortgage originators and secondary market issuers to adopt policies and controls so that automated valuation models used to value a consumer's principal dwelling maintain a high level of confidence in the estimates, protect data integrity, avoid conflicts of interest, are tested by random sample review and comply with nondiscrimination law.
Controls to put in place
- A qualified valuer signs off on every high value, high risk or low confidence automated valuation before it is relied on
- An independently reviewed accuracy and fairness study by property type and area, refreshed on a set cycle
Frequently asked questions
- Can AI fully replace a professional valuer?
- It depends on the deployment. Riverside County routes properties outside its configured AVM variance thresholds to a person, and the Valuation Office Agency routes the batches its model scores as least reliable or closest to a Council Tax band boundary to a valuer; RICS is explicit that automated valuation models should support, not replace, the valuation process for lending, legal or investment decisions. Fannie Mae's Value Acceptance goes further and replaces the appraisal entirely for eligible loans, within set loan to value and loan type limits, rather than routing anything to a human valuer.
- How much do lenders and governments save with automated valuation?
- Fannie Mae estimates that appraisal alternatives such as Value Acceptance saved US mortgage borrowers more than $2.5 billion since early 2020, though that figure also covers Value Acceptance + Property Data, which uses a third party data collector rather than a fully automated valuation. The UK Valuation Office Agency's own early estimate for its Wales Council Tax revaluation model is a reduction in the cost of a revaluation by one third compared with a fully manual valuation, an estimate for a revaluation planned for 2028 that has not happened yet.
- How accurate does an automated valuation need to be?
- There is no single number. C3 AI reports a 40% improvement in model accuracy at Riverside County, over the county's previous linear regression models, demonstrating the ability to directly enroll up to 97% of all property sales. Independent testing against a recognized mass appraisal standard, not a vendor's own figure alone, is what a buyer should ask for.
- Is an automated valuation high risk under the EU AI Act?
- An automated valuation model is not itself listed in Annex III; it values the collateral, not the borrower. It becomes relevant to Annex III point 5(b), creditworthiness assessment of natural persons, only where its output is built into a separate system that assesses the borrower, which depends on how the lender designs that system, not on the valuation model alone.
How to cite this page
Blits.ai AI Use Case Library, "AI support for property valuation and appraisal", last verified 28 September 2026, https://www.blits.ai/ai-use-cases/property-valuation-support. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 28 September 2026: First published