AI Tenant Screening and Property Valuations: EU AI Act Obligations
Letting agencies now score applicants in seconds — and if that score touches creditworthiness, the EU AI Act treats it like a bank decision. Where the Annex III line falls for screening, valuations and rental pricing, and how to fix a fully automated rejection.
A rental application filtered by an algorithm looks nothing like a loan decision — until you notice that the model’s heaviest input is the applicant’s creditworthiness. That single fact drags tenant screening into the same Annex III category as bank lending, with consequences that most proptech vendors and letting agencies have not priced in: provider duties for the software, deployer duties for the agency, and a hard GDPR Article 22 problem whenever a human never sees the rejection.
Valuations, rent-pricing and maintenance triage sit at calmer points on the map. Knowing which is which is the entire game.
Where tenant screening lands in Annex III
Annex III point 5(b) covers AI systems used to evaluate the creditworthiness of natural persons — the category the Act inherited from consumer-credit regulation. Screening an apartment applicant by:
- Credit-bureau scores or affordability models — this is creditworthiness evaluation. High-risk.
- Income-to-rent models fed by payslips and bank data — creditworthiness by another name. High-risk.
- Rental-history and eviction-risk prediction — sits close enough to point 5(b) that defensible teams treat it as in scope, especially when it drives the decision rather than informing it.
- Identity or right-to-rent checks performed by a human — generally outside the AI Act’s high-risk scope; ordinary GDPR processing applies.
- Reference-reading automation with a person deciding — support tooling, lower exposure; document the human’s authority anyway.
The derogation question. Article 6(3) lets certain Annex III systems escape high-risk treatment where they only improve the result of a pre-existing human decision, do not replace it, and pose no risk to health, safety or fundamental rights — plus profiling concerns. Screening pipelines love to claim this exemption. Two things blunt it: the derogation does not apply where the system profiles natural persons (scoring applicants is profiling in the GDPR sense), and it never applies where an Annex III area has been carved out by the Annex III notes themselves. Treat the derogation as an argument to document, not a switch to flip.
The valuation and pricing end of the market
Not every proptech model is a scoring engine:
| Use case | AI Act position | Why |
|---|---|---|
| Automated valuation model (AVM) for mortgages | High-risk where the valuation feeds lending decisions (creditworthiness chain) | Annex III point 5(b) read-across |
| Rental price estimation for landlords | Generally minimal risk | No evaluation of a natural person |
| Applicant credit scoring | High-risk | Annex III point 5(b), explicit |
| Eviction / rental-arrears risk on sitting tenants | High-risk territory | Evaluates a person’s financial reliability |
| Maintenance triage and viewing scheduling | Minimal risk | Administrative support |
| Dynamic rent pricing across a portfolio | Minimal risk under the AI Act — but watch national tenancy law and unfair-practices rules | The Act regulates persons, not prices |
The AI Act looks at people, not property. Models that price buildings stay in the calm zone; models that rank applicants or sitting tenants walk into Annex III. Vendor marketing blurs this line deliberately — “AI-powered screening” in a pitch deck does not tell you which side you are on; the input features do.
GDPR Article 22: the sharper edge
For most agencies the binding constraint arrives before the AI Act does. If a rental application is decided solely by automated processing with legal or similarly significant effects, the applicant has the right not to be subject to that decision under GDPR Article 22, with narrow exceptions (contract necessity, authorised law) — and obtaining “meaningful human involvement” is the usual compliance route. Regulators have repeatedly held that a human rubber-stamping the model output is not meaningful involvement. Concrete fixes:
- Insert a genuine review step for every adverse decision — a named person who reads the file, can override, and records why. Timebox it so it cannot become a rubber stamp.
- Tell applicants up front — plain-language notice that scoring is used, what it considers, and how to contest it; this also feeds the AI Act’s deployer information duties where they apply.
- Log the overrides. If your override rate is 0.02%, you have documented your way out of “meaningful” — audit trails cut both ways.
Deployer and provider duties when screening is high-risk
When screening sits in Annex III point 5(b), obligations split along the value chain:
- Providers (software vendors) carry conformity assessment, technical documentation (Annex IV), registration, and the duty to supply instructions that enable deployer compliance.
- Deployers (agencies, property managers) owe the Article 26 set: assign human oversight with competence and authority, ensure input data is relevant and representative enough for the purpose, inform workers and affected applicants, keep logs of use, and — before first use at work — inform workers’ representatives and offer consultation where national law or Article 26(7) requires it.
- Both must watch for fundamental-rights flashpoints: housing decisions affecting vulnerable applicants sit exactly where supervision authorities look first.
For mid-market agencies the practical order: classify honestly, fix the Article 22 hole, put the register entry and log retention in place, then evidence the vendor’s conformity position contractually.
What to demand from your screening vendor
- Their Annex III analysis for point 5(b) — and where the system sits if they claim an Article 6(3) derogation.
- EU database registration status, if they are a provider of a high-risk system.
- Input-data documentation: which bureau feeds, what features, how often retrained, how bias is checked — this becomes your input-data duty as deployer.
- Meaningful-human-review support in the product: workflow flags, override capture, reason codes. If the tool cannot record a human decision, it cannot help you satisfy Article 22 or Article 26.
- Incident and change notification: model updates that shift decision boundaries are your re-assessment trigger.
How to check your own position in three minutes
The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves — mapped to the dates that apply. Confirm your email and the full report unlocks.
Frequently asked questions
Is every tenant screening tool high-risk under the AI Act?
No — only the parts that evaluate an applicant’s creditworthiness. Identity checks, reference handling by a person, and administrative automation sit outside Annex III point 5(b). Map the actual decision path: if a score of the person’s financial reliability gates the outcome, that step is high-risk; if the model only sorts viewing slots, it is not.
We always have an agent approve the final decision — does Article 22 apply?
Only if the human’s involvement is real. Supervisory authorities look at whether the reviewer could realistically change the outcome, had access to the underlying information, and recorded their reasoning. A workflow where the agent sees only “recommended: reject” and clicks OK will be treated as automated decision-making — build the review step so it can and does overturn the model.
Do valuations and rent-pricing algorithms need conformity assessment?
Generally not. The AI Act targets systems that evaluate people; models pricing property do not enter Annex III point 5(b) merely because they are AI. Mortgage-linked valuations are the exception, because the number ends up as an input to a creditworthiness decision — classify the chain, not the isolated model.
What does an agency actually have to do as a deployer?
Assign and document human oversight, check input data quality before go-live, inform affected applicants, retain logs of system use, register the high-risk use where required, and consult workers’ representatives before workplace deployment. Most of this collapses into artefacts you can assemble in weeks if the vendor supplies the right documentation.
Can we claim the Article 6(3) derogation for our screening flow?
Possibly, but not if the system profiles applicants — and screening by definition scores people against behavioural and financial patterns. Document the derogation test honestly (improves a human decision, does not replace it, no fundamental-rights risk, no profiling), and be prepared for a supervisory authority to disagree on the profiling point.
What is the fastest fix if we are already live with automated rejections?
Insert genuine human review of every adverse decision with recorded reasons, add the applicant-facing transparency notice, start logging overrides, and open the vendor documentation conversation for the Annex III position. None of that requires pausing operations — and it closes the GDPR exposure that is legally live today, well before the AI Act’s high-risk dates arrive.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.
Keep reading
AI in healthcare under the EU AI Act: clinical decision support and high-risk duties
11 min
Risk & ClassificationThe Annex III High-Risk AI Systems List: All Eight Use Cases Explained with Examples
11 min
Risk & ClassificationThe Eight Prohibited AI Practices Under Article 5: A Compliance Checklist
10 min