Can an AI tool screen a rental applicant?

Two separate bodies of rules apply and they ask different questions. The housing rules ask about the outcome: a policy or practice that disproportionately excludes people with a protected characteristic can be unlawful whether or not anyone intended it, and a housing provider is answerable for a decision made on its behalf. The consumer reporting rules ask about the process: where a decision is taken on information in a consumer report, the applicant has to be told, told who supplied the information, and told they can dispute it. A tool can satisfy neither and a vendor can satisfy neither on your behalf.

The regulator's guidance, issued in 2024, settled the first question for automated systems specifically, addressing tenant screening and housing advertising that rely on algorithms. It confirmed what the general rule already implied — that housing providers and the screening companies they use can both be answerable — and it removed the argument that a purchased score is somebody else's product.

The screening inputs that cause most of the trouble are the ones that look neutral and are not. Eviction filings that were dismissed or resolved, arrest records without conviction, credit history in categories where access to credit itself varies, income multiples applied uniformly to applicants with housing subsidies, and gaps in rental history that reflect circumstances rather than conduct. Each has been the subject of enforcement or litigation in its human form, and a model trained on historical decisions inherits every one of them without anyone choosing it.

Which is why the design question is not accuracy but auditability. A screening arrangement you can defend is one where you can state which factors were considered, what weight each carried, what the threshold was, and how outcomes distribute across applicant groups. A model returning a single risk score cannot supply any of that, and the position that follows — we do not know why it declined this applicant — is worse than a bad reason.

The adverse action step is the one that is most often simply missing. A decline based in whole or in part on a consumer report requires notice with the source named and the applicant's dispute right stated, and it is a procedural requirement that does not care how good the underlying decision was. Automated pipelines routinely decline and move on, because nobody built the notice into the flow, and a missing notice is a complete and easily proven claim regardless of whether the decision itself was sound.

The version of this that works keeps the model on the administrative half. Reading applications, extracting stated income and employment, checking that documents are complete, ordering the reports, flagging inconsistencies for a person, and organising the file so a human decision takes minutes rather than an afternoon. Every one of those is genuine time saved and none of them decides anything. The moment the system produces a rank or a recommendation, a person following it has adopted it, and the record will show they did.