Real estate planning teams stopped debating whether to use artificial intelligence some time ago. The live question in 2026 is narrower and more awkward: which decisions a model is allowed near, now that regulators on both sides of the Atlantic have written rules about exactly that.
Four things to hold on to
- Site screening and feasibility studies are where the time savings are real and the risk is contained.
- Since 1 October 2025, US mortgage originators and secondary market issuers using automated valuation models must meet five interagency quality control standards.
- The EU deferred most high-risk AI Act obligations to 2 December 2027, but transparency and AI literacy duties were not postponed.
- Model quality is capped by the quality of cadastral, permit and floor area data, which nobody has cleaned for you.
Where models actually sit on the planning desk
Four tasks account for most serious use. They differ far more in their failure modes than in their technology.
| Task | What it does well | How it fails |
|---|---|---|
| Site screening | Filters hundreds of parcels against zoning, access, flood and ownership criteria | Silently drops sites whose records are incomplete rather than unsuitable |
| Massing and feasibility | Generates many compliant envelopes and unit mixes quickly for comparison | Optimises the constraints it was given, not the ones a planning officer will raise |
| Valuation | Produces fast, consistent estimates across a large portfolio | Reproduces the patterns in its training data, including the ones nobody wants reproduced |
| Demand and absorption forecasting | Spots turning points in listing, permit and mobility data earlier than a quarterly report | Extrapolates confidently through exactly the conditions it has never seen |
The pattern is consistent. Analysis that feeds a human judgement is low risk. Output that becomes the judgement is not, and the regulatory activity of the past two years has landed precisely on the second category.
The valuation rule that came into force in October 2025
In the United States, six federal agencies including the OCC, the Federal Reserve, the FDIC, the NCUA, the CFPB and the FHFA adopted a final rule on quality control standards for automated valuation models, which took effect on 1 October 2025. It applies to mortgage originators and secondary market issuers using an AVM to determine the collateral value of a mortgage secured by a consumer’s principal dwelling.
The rule requires firms to adopt policies and controls designed to deliver a high level of confidence in the estimates, protect against data manipulation, avoid conflicts of interest, and carry out random sample testing and reviews. To those four statutory standards, the agencies added a fifth of their own: compliance with applicable nondiscrimination laws. That addition is the substantive one. It puts the burden of demonstrating that a model does not produce discriminatory outcomes on the institution using it, not on the vendor that built it.
Europe moved its deadline, but not all of it
Anyone who planned around 2 August 2026 as the compliance date for high-risk AI systems needs to re-read the calendar. The Digital Omnibus on AI was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026, days before that deadline. It defers the obligations for standalone high-risk systems listed in Annex III to 2 December 2027, and for AI embedded in products already covered by EU product safety law to 2 August 2028.
Two caveats matter for property teams. The deferral did not touch the transparency obligations or the AI literacy duty, which remain on their original timetable. And Annex III includes systems used to evaluate the creditworthiness of individuals, so a tool sitting in a residential lending chain may be in scope even when a pure asset valuation tool is not. Classification depends on the intended purpose of the system, which is a question for counsel rather than for a procurement form.
Both regimes point at the same practical housekeeping, whichever market a firm operates in. Keep a plain register of the models in use, what each one decides or informs, which data it was trained or calibrated on, who reviewed its output and when it was last tested. Firms that already maintain one have found the compliance work to be mostly documentation of things they were doing anyway. Firms that do not tend to discover that nobody can say with certainty how many tools are in production, which is an uncomfortable answer to give a supervisor.
The unglamorous constraint: the data underneath
Every model on a planning desk inherits the quality of the public record it was trained on. Cadastral boundaries carry historic surveying error. Permit databases record what was applied for rather than what was built. Floor areas are measured to different standards in different markets, and sometimes in the same market by different agents. None of this is visible in the output, which arrives with the same confident formatting whether the underlying parcel record is immaculate or forty years stale.
The practical discipline is unchanged from before any of this existed: spot-check the inputs on a sample of assets before trusting the conclusions across a portfolio, and keep the sample large enough to be uncomfortable.
What stays with a person
Three decisions are worth ring-fencing regardless of how good the tools become. The first is anything that determines a price offered to a consumer, for the regulatory reasons above. The second is the reading of a planning authority, which turns on precedent, politics and personalities that are not in any data set. The third is the go or no-go on a scheme, because accountability for that decision has to sit with someone who can be questioned about it.
Once a scheme is built, the same analytical machinery becomes considerably less contentious, since it is then working on the asset’s own operating data rather than on inferences about people. That is the argument for predictive maintenance across a property portfolio, where the feedback loop is short and the errors are cheap.
Used this way, artificial intelligence changes the economics of the early stages of a development without changing who is answerable for the outcome. That distinction is doing a lot of work in the current regulations, and it is a reasonable one to build a process around.
This article is general information about tools and current regulation, not investment, valuation or legal advice. Requirements differ by jurisdiction and change quickly; take professional advice on any specific transaction.
Interested in the scale above a single site?
The same techniques applied at city scale raise a different set of questions about data, consent and public accountability.
