Property managers can screen tenant applications fairly with AI by fixing the criteria in writing first, then using it only for consistent admin: checking applications are complete, matching income evidence to the stated figure, summarising references, and producing the same summary for every applicant. A person makes every accept-or-decline decision and records the reason. AI never scores applicants or reads social media.
Fairness is the hard part because unfairness rarely looks like prejudice. It looks like a sensible-sounding rule, applied by a tool nobody questions. In November 2024 a court approved a $2.3 million settlement in a class action against tenant-screening company SafeRent, brought over claims that its scores disadvantaged applicants who paid part of their rent with housing vouchers. As part of the settlement, the company agreed for five years to stop giving voucher holders a score or an accept/decline recommendation. The score didn't ask anyone's ethnicity. The claim was that it leaned on credit history, including debts unrelated to renting, while ignoring the voucher that guaranteed much of the rent, and that this produced lower scores for Black and Hispanic applicants. The same risk sits in any AI tool that learns from past data.
Write the criteria before the first application arrives
Criteria written after you've seen the applicants aren't criteria; they're justifications. Put them on one page, apply them to every application for a property, and change them only by a dated decision. An illustrative filled-in sheet for a two-bedroom flat at $1,200 a month:
| Criterion | Standard for this property | Evidence accepted |
|---|---|---|
| Affordability | Gross household income from any lawful source of at least 30 times the monthly rent ($36,000 a year), or a guarantor earning at least $43,200 | Payslips, bank statements, pension or benefit award letters, accounts for self-employed applicants |
| Rental history | No unresolved rent arrears with a previous landlord in the last 3 years | Landlord reference; if no previous tenancy, a character or employer reference instead |
| Identity | Identity confirmed for every adult occupant | As your legal checks require |
| Occupancy | Maximum 4 people, in line with the property's licence and size | Application form |
| Move-in date | Within 6 weeks of the property being available | Application form |
| Deposit and first month | Payable before keys | Payment |
Notice "from any lawful source": pension, benefits and self-employed income count the same as a salary. Notice too what isn't on the sheet: a minimum credit score, "stable employment", a job title, a tidy address history. Each of those can exclude groups your equality and anti-discrimination law protects, often without anyone intending it.
What AI may do in screening, and what it must not
| Safe uses (with a person checking) | Uses to avoid |
|---|---|
| Checking every required document is present | Scoring or ranking applicants |
| Extracting income figures into a standard summary | Recommending accept or decline |
| Comparing payslips with bank statements for consistency | Reading social media or searching applicants online |
| Summarising references against the criteria | Judging "fit", personality or tone of an application |
| Drafting requests for missing documents | Inferring age, nationality, religion, health or family plans |
| Drafting decision letters from the reason a person recorded | Choosing between two applicants who both meet the criteria |
Where two applicants both meet the criteria, use a rule you've written down in advance, such as first complete application received, rather than a judgement.
Define "complete" precisely, or the tie-break becomes its own argument. Take two illustrative applicants for the same flat. One applies at 9am on Monday, but their landlord reference arrives on Wednesday; the other applies at 3pm on Monday with everything attached. Under "first complete application", the second applicant is first, and the log should show both timestamps. Have the AI's completeness check record when the last required document arrived, not when the form was submitted, so the rule is applied from the record rather than from anyone's memory of who seemed keener.
The broader case for keeping decisions like these away from AI is in AI bias in small business decisions.
One summary format for every applicant
The fairness benefit of AI, used properly, is consistency: every applicant is summarised in the same structure against the same criteria, however their documents were laid out. The prompt:
You are preparing a factual summary of a tenancy application for a
property manager. Use ONLY the documents provided.
Property criteria: [paste criteria sheet]
Produce:
1. A table: each criterion, the evidence found, "met", "not met" or
"evidence missing".
2. Monthly income figures you extracted, each with its source document
and page.
3. Inconsistencies between documents, stated factually.
4. A list of documents requested but not supplied.
Do not give an overall recommendation. Do not comment on the
applicant's character, job, background or writing. Refer to the
applicant as "Applicant [reference]".Output for one applicant (illustrative):
Affordability: MET. Salary $2,150/month (payslips Jun-Aug) plus
pension $1,040/month (award letter). Total $38,280/year.
Rental history: MET. Previous landlord reference: no arrears.
Inconsistencies: August payslip net pay $1,702; bank statement shows
$1,702 received 28 Aug. None found.
Note: applicant recently changed jobs (new employer since May), which
may indicate instability.The table is what you want. The final note is what the prompt forbade and what the model added anyway: "may indicate instability" is a judgement that isn't in the criteria, and job changes aren't a criterion here. Delete it, and if it keeps appearing, add "do not add notes or observations beyond items 1 to 4" to the prompt. Check the arithmetic too: $2,150 plus $1,040 is $3,190 a month, or $38,280 a year, so this one is right, but it's the kind of figure a model occasionally gets wrong.
References need the same discipline, because summarising is where a model's own judgement creeps back in. An illustrative line from a previous landlord: "Rent was paid late on two occasions in 2022, both cleared within a week. No arrears at the end of the tenancy and we would let to them again." One unguided summary turned that into "Rental history: some history of late payment." Against this property's criterion, no unresolved arrears in the last 3 years, that's misleading: the criterion is met. Add to the prompt: "For each reference, quote the sentence relevant to each criterion, then state met or not met against the criterion's exact wording." The quoted sentence lets whoever decides check the call in seconds.
Before the summary reaches whoever decides, strip what the decision doesn't need: date of birth, photos, nationality details from identity documents, and names where your process allows. Identity checks happen separately. Redacting personal data from documents with AI covers how to do that without missing fields.
Covering letters are the usual leak. Applicants volunteer things: "I'm 63 and recently retired", "we're expecting our second child in the spring", "I use a wheelchair, so step-free access matters". Test your process by planting sentences like these in a dummy application and checking whether the summary repeats them. It shouldn't, and the prompt's "do not comment on background" line usually holds, but check rather than assume. The wheelchair sentence needs its own route: a request connected to disability may need a response under your equality law, so whoever handles the raw application should pass it to the person who deals with adjustments, outside the screening summary. Ask your legal adviser what your obligations are.
Checking income and documents without guessing
AI can compare figures across documents much faster than a person, and it's good at noticing that a payslip's net pay doesn't match the deposit on the bank statement. Treat every mismatch as a question to ask, not a sign of fraud. People are paid late, employers make corrections, and joint accounts muddle things.
In practice it goes like this. An applicant's payslip shows net pay of $2,410 but the bank statement shows $2,160 arriving. The AI flags it. The manager asks. The applicant explains that $250 goes straight to a savings account by salary sacrifice, and sends the confirmation. Five minutes, no harm done. If the flag had triggered an automatic decline, a qualified applicant would have been refused for having a savings habit.
Joint applications and guarantors trip up summaries more quietly. In an illustrative case, two sharers applied together, earning $19,000 and $18,500. The criterion is household income, so together they have $37,500 against the $36,000 standard: met. The first AI summary listed only the lead applicant's payslips and marked affordability "not met", because the second sharer's documents sat in a separate upload. Unchecked, the refusal letter would have quoted a figure that was simply wrong. Two fixes: put every applicant's and guarantor's documents into the same request, and add to the prompt "list each adult applicant and any guarantor separately, then give the combined household figure". A guarantor is tested against their own standard ($43,200 here) and never added to household income.
The decision, the reason and the letter
A person reads the summary, applies the criteria, and records the decision with one reason tied to a criterion. The AI can then draft the letter from that recorded reason. Two versions of the same refusal:
Unguided AI draft: "After careful consideration, we have decided not to proceed with your application as we did not feel it was the right fit for this property."
Draft from the recorded reason: "Thank you for applying for the flat. We're unable to offer you the tenancy because the income evidence provided ($29,400 a year) is below this property's affordability standard of $36,000, and no guarantor was named. If your circumstances change, or you can add a guarantor, you're welcome to apply for this or another property. If you think we've misread your documents, reply to this email and a different member of the team will review your application."
"Not the right fit" is the phrase that makes a decision look discriminatory, because it could mean anything. The second letter names the criterion, the figure and the route to a review. That review route matters legally as well as ethically: under data-protection law such as the GDPR, people have rights around decisions made solely by automated means that significantly affect them, and a refused home plausibly qualifies. A person deciding, with a reason on file, keeps you on the right side of that.
When a neutral rule turns out to be unfair
Most unfair screening comes from rules that sounded reasonable. A realistic example: a firm's old criteria said "income from employment of at least 30 times the rent". It excluded retired applicants living on pensions, disabled applicants receiving benefits, and self-employed people with a thin first year of accounts, none of whom were more likely to fall behind on rent. Nobody intended it. It surfaced only when someone counted the declines by income type.
Other rules to test: a minimum credit score (thin files are common among young people, recent arrivals and people who've been through divorce); "no gaps in address history"; requiring a previous landlord reference, which first-time renters can't provide; and anything that relies on how well an application is written. The same pattern shows up in recruitment, and whether AI can screen CVs fairly walks through the equivalent tests.
If you're buying a screening product
Ask before signing: what data the tool uses, whether it produces a score or a recommendation, how it has been tested for bias, whether you can see the reasons behind each result, whether you can switch off automated recommendations, and how applicants challenge a result. The general version of that checklist is in questions to ask before buying AI that touches client data.
If you let homes to people in the EU, note that a tool which evaluates individuals' creditworthiness or produces a credit score falls within the EU AI Act's high-risk category, with those obligations now applying from 2 December 2027. Take advice on whether a particular product is caught.
A monthly fairness check you can actually run
Keep one log row per application: property, date, criteria met, decision, reason, decided by. Once a month, spend 30 minutes on three questions: do declines cluster by income type, guarantor use or any other pattern; did every decline cite a criterion; and does a second person agree with a random sample of five decisions? An illustrative month:
| Income type | Applications | Declined | Main reason |
|---|---|---|---|
| Employment | 26 | 5 | Below affordability standard |
| Self-employed | 6 | 3 | Accounts not supplied |
| Pension or benefits | 7 | 1 | Below affordability standard |
Half of self-employed applicants declined, all for missing accounts, is a pattern worth a look. It may be fine, or it may mean your document request is unclear for self-employed people. That month, the manager rewrote the request to say exactly which accounts were acceptable, and the next month's self-employed declines fell to one.
The rewrite is worth seeing, because vague document requests are where many unfair-looking declines start. Before: "Please provide proof of income." After: "If you're self-employed, please send either your last two years' accounts prepared by an accountant, or your last two years' tax returns with the tax authority's calculation. If you've traded for less than two years, send what you have plus bank statements for the last six months. If you can't provide any of these, reply and tell us what you do have, and we'll say whether it can work." The AI drafted it in a minute from the criteria sheet; the manager checked that every document it named was one the firm would actually accept.
The second-person sample earns its 10 minutes too. In the same illustrative month, the reviewer disagreed with one of five decisions: a decline for rental history where the reference showed arrears that had been cleared under an agreed repayment plan two years earlier. The criterion says "unresolved" arrears, so it was met. The firm contacted the applicant, apologised and offered a viewing of the next suitable vacancy, and the criteria sheet gained one line defining "unresolved" as a balance still outstanding on the date of the application.
The time saved across a 350-home portfolio
An illustrative property management firm looks after 350 homes and handles about 40 applications a month for around 12 vacancies. Before AI, a manager spent roughly 45 minutes per application chasing documents, reading payslips and writing up notes. With the standard summary and document-request drafts, that falls to about 20 minutes, around 16 hours a month back, plus a few minutes per decision letter.
The less visible gain is that every applicant now gets the same process, and the firm can show it. When a landlord client asks why an applicant was refused, or an applicant challenges a decision, the answer is a criteria sheet, a summary, a recorded reason and a log. Tenant queries once they've moved in are a separate workload, covered in AI for letting agents handling tenant queries.
Screening tools and fairness: questions managers ask
Can we use a screening service that gives each applicant a score?
You can use one as an input, if you know what goes into the score, can see the reasons behind it and never decline automatically on the number alone. Ask the vendor how it tests for bias against groups such as people on benefits or pensions, and what an applicant can do to challenge a result. If they can't explain the score to you, you can't explain your decision to the applicant.
Is it fair to look at applicants' social media?
No. It's inconsistent, because some applicants have public profiles and others don't, and it exposes you to exactly the information you must not use: religion, ethnicity, sexuality, health, family plans. Once you've seen it, you can't show you ignored it. Keep screening to the documents your written criteria ask for, and never ask an AI tool to research applicants online.
Do we have to tell applicants that AI is involved?
Your privacy notice should say what data you collect, what software processes it and why. Beyond that, it's good practice to tell applicants how applications are assessed and that a person makes every decision. Where a decision is made solely by automated means and significantly affects someone, the GDPR and similar data-protection laws give them extra rights, including human review, which is one more reason to keep a person deciding.
Further reads
- Is AI CV Screening Fair? Bias Checks Every Recruiter Should Run — The bias checks recruiters run, which translate well to tenant screening.
- AI Governance for a Small Business: Who Decides, Approves, Checks — Decide who approves criteria changes and who audits decisions.
- A Simple AI Risk Register for Small Businesses (With Template) — Record screening risks and the controls you've put in place.
- How Much of a Property Manager's Week Can AI Take Over? — Where screening sits among the rest of a manager's week.
- AI Compliance Checklist for Small Businesses: What Applies to You — A wider compliance check before any AI tool goes live.
- AI Landlord Updates and Owner Statements in Minutes — Report screening outcomes to landlords clearly.
- AI Inventory and Check-Out Reports: A Letting Agent's Guide — How AI drafts inventory and check-out reports from photos or video, where its descriptions go wrong, and how to keep the report defensible in a deposit dispute.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: court-approved settlement in Louis v. SafeRent Solutions (November 2024), as reported by class counsel; EU AI Act Annex III and implementation timeline; GDPR Article 22.