Is AI Tenant Screening Reliable? Answering the Experts' Criticism

Artificial intelligence has moved quickly into tenant referencing, and not everyone in the industry is convinced. Some of the most experienced figures in UK referencing have warned — with good reason — that reading a few bank statements is a weak basis on which to bet a year's rent. If you are a landlord weighing up whether to trust an AI screening report, that scepticism deserves a serious answer rather than a marketing brush-off.

So let us take the criticism head-on. The experts are right about the thing they are actually criticising. Where they are wrong is in assuming that thing is what modern AI pattern analysis does.

The Criticism Is Fair — For Naive Checks

The strongest version of the sceptical case comes from people who reference tenants for a living, and it is worth stating properly before answering it.

In The Negotiator's feature on the cost of checking, Andy Halstead, founder and CEO of Let Alliance, makes a specific and genuinely insightful point. The danger, he argues, is not the tenant who already looks bad — it is the one who looks fine today but whose finances are quietly turning. If credit use is climbing while repayments fall in relation to what is owed, an experienced analyst can conclude that a tenant is only "a few weeks away from disaster". And, he adds, "you have to have people analysing that."

Graham Sandley, managing director of Diligent, puts the limitation of thin data bluntly: you are "always trying to make a future projection," and "relying on a few bank statements ... isn't going to work."

Both are correct — about the thing they are describing. A crude check that only shows money in and money out is a snapshot. It tells you the balance cleared last month; it tells you nothing about the direction of travel. A snapshot cannot distinguish a stable tenant from one whose credit reliance is accelerating toward a cliff edge. If that is all "AI screening" meant, the sceptics would be right to dismiss it. This is the same structural weakness we describe in why credit checks alone aren't enough: a single moment in time is not a trajectory.

The Turn: That's a Description of Trajectory Analysis

Here is the crucial point. Read Halstead's warning again — rising credit use, falling repayments, a tenant weeks from disaster. That is not a description of something only a human can do. It is a description of a time-series pattern. It is, almost word for word, the definition of what automated affordability analysis is built to compute.

The critique lands squarely on naive open banking that dumps a raw transaction feed on screen — the "just shows money in and out" version. It does not land on analysis that examines patterns across months of real financial behaviour: whether income is regular, whether credit dependency is trending up, whether repayments are keeping pace with balances, whether the buffer is thinning. The very trend the expert says requires a human is exactly the trend a machine can trace across every transaction — and trace the same way, every time.

In other words, the sceptics have accurately diagnosed the weakness of a snapshot, and then attributed it to AI. But the modern answer to a snapshot is not a human analyst. It is a trajectory.

See the direction of travel, not just today's balance. LetSorted analyses patterns across months of real financial behaviour — income stability, debt trends, and affordability against the rent asked — and hands you a clear grade to decide on. Screen your next tenant →

What AI Actually Does Better (Stated Honestly)

None of this means AI "sees everything" or that human expertise is worthless. It means AI is stronger in the specific places where human manual analysis is objectively weak. Four of them matter most.

1. Volume — every transaction, not a skim

A human analyst under time pressure skims. Faced with several months of statements across multiple accounts, no one reads every line — they sample, and they form an impression. Automated analysis reads all of it: every transaction, every recurring debit, every credit line, across the whole period. The debt trajectory Halstead describes is easy to miss in a skim and hard to miss in a full pass. Coverage is where machines have an unfair advantage.

2. Trajectory and pattern — computed, not eyeballed

Spotting that credit use is rising relative to repayments, that income arrives irregularly, that liquidity is thinning month over month — these are multi-variable trends. A person estimates them by eye; a system computes them consistently. This is precisely the "few weeks away from disaster" signal, quantified rather than intuited. For how this sits alongside other methods, see our comparison of screening approaches.

3. Consistency — no human factor

This is the quietest advantage and arguably the most important. The quality of a manual reference depends on which clerk is at the desk that day: their competence, their training, whether it is 9am or the end of a long shift, and their unconscious assumptions about an applicant. Two identical files can get two different answers. A model applies the same criteria to everyone. That consistency is not only more reliable — it materially reduces the risk of the inconsistent, subjective judgements that create discrimination exposure under the Equality Act 2010.

4. Speed and inclusivity

Analysis that would take an analyst hours runs in minutes, which matters in a competitive market. And it works for applicants that traditional referencing structurally fails: recent migrants with no UK credit file, high earners whose income and savings sit in overseas accounts, the self-employed with irregular income, and anyone with a thin credit history. Assessing real transaction data rather than a bureau score widens the pool fairly rather than excluding people for lacking a UK borrowing record.

What AI Does Not Do (The Honest Limits)

A fair answer to the sceptics has to concede the ground where they are still right — because overclaiming here would repeat exactly the mistake they are warning against.

• It does not replace your judgement. Well-designed screening is decision-support: it surfaces the evidence and grades the risk; the landlord makes the call. Under the ICO's guidance on automated decision-making, applicants also retain the right to request a human review, so a human stays in the loop by design.
• It is only as good as the data it sees. If an applicant submits one of three accounts, the analysis reflects one of three accounts. Scope limits accuracy — Sandley's warning about thin data holds whenever the inputs are thin.
• It does not capture tenancy behaviour. Whether someone paid rent on time and left the last property in good order comes from a previous-landlord reference, not a bank feed. AI complements references; it does not retire them.
• It is not a crystal ball. A trajectory is a strong signal, not a guarantee. Circumstances change after any check. The honest claim is that AI reads the present direction of travel more thoroughly and consistently than a manual skim — not that it predicts the future.

Position it correctly and the picture is clear: AI removes the weaknesses of manual analysis — limited coverage, inconsistency, fatigue, unconscious bias — rather than pretending those weaknesses never mattered.

The Verdict: Reliable Where It Counts

So, is AI tenant screening reliable? For the work it is actually doing — reading every transaction, computing the debt-and-income trajectory, and applying identical criteria to every applicant — it is more reliable than the manual alternative, not less. The experts were right that a snapshot is not enough and that spotting a deteriorating trend takes real analytical work. They were simply describing the previous generation of checks.

The best screening process still combines signals: a credit check for historical court records, AI statement analysis for current trajectory and affordability, references for tenancy behaviour, and a Right to Rent check for legal compliance. For the full workflow, see the complete guide to screening tenants in the UK. Used that way — as the objective, consistent layer rather than the whole decision — AI does the thing the sceptics said needed doing, and does it to every applicant, every time.

Frequently Asked Questions

Is AI tenant screening reliable?

For the things it is objectively good at — reviewing every transaction across months, detecting income and debt trajectories, and applying the same criteria to every applicant — AI is more consistent than manual analysis, which varies with whichever clerk is at the desk. It is not infallible: it depends on the quality of the statements provided and does not replace the landlord's final judgement or a previous-landlord reference. Used as decision-support, it is reliable where reliability matters most.

Can AI replace human analysts in tenant referencing?

Not entirely, and it shouldn't. The critique that spotting a subtle debt trajectory "requires a human" was accurate for naive checks that only show money in and money out. Modern AI pattern analysis computes those trajectories directly and consistently, but the landlord still makes the decision and applicants retain the right to a human review under UK GDPR. AI removes the weaknesses of manual analysis — fatigue, inconsistency, unconscious bias — rather than removing the human from the loop.

How accurate is bank statement analysis for screening tenants?

Accuracy depends on scope and method. A single snapshot of a few statements is a weak basis for a future projection, which is a fair criticism. Analysing patterns across several months of real transactions — income regularity, rising credit use, falling repayments, liquidity — is a far stronger signal, because it captures the direction of travel rather than a single moment. It is most accurate when combined with a credit check and references.

This article is for informational purposes only and does not constitute legal advice. Quotations are attributed to their original speakers and sources and are reproduced here for commentary and criticism.