A letting agent can use AI to compare referencing documents with each other, check the arithmetic on payslips and statements, and flag signs of tampering such as odd file properties or mismatched dates. The AI only raises flags. A person confirms each one by calling the employer or landlord, and nobody is rejected on a flag alone.
Fake payslips and bank statements are now easy to produce, so a quick visual check no longer works. But the same tools that make fakes also make a decent first-pass reviewer: they are patient with columns of numbers, and they never tire at the thirtieth application of the week. What they cannot do is know what is true. This checklist is built around that division.
The rule that comes before every check
Write this at the top of your referencing procedure: a flag is a question, never a verdict. Every item below produces a flag, a note to a person, and a next step that is a phone call or a request for another document. None produces a rejection. That protects honest applicants, who often have odd-looking documents for ordinary reasons (a freelancer with a hand-made invoice, a new employer's basic payslip template, a bank statement printed from an app), and it protects you from acting on a model's confident mistake.
Also decide, in advance, who signs off. A named senior member of staff, not the person who ran the check, should record the decision, the reason and the evidence. If an application is declined, the reason must be one you can explain to the applicant in plain words and defend, so ask a qualified adviser what equality and data-protection rules require in your business before you use any automated check at scale.
What to put in front of the AI, and what to keep out
Referencing packs hold sensitive data: identity documents, account details, employer names, addresses. Before using a general chatbot, check the plan's data terms, and prefer a tool your agency has a contract for. Redact what the check does not need: account numbers, identity numbers, dates of birth, home addresses. The check is about consistency between documents, which works fine with "Employer A" and "Applicant 1".
Keep a short log for each application: which documents were checked, by which tool, on what date, what the flags were and who made the decision. If a dispute arises, that log is your evidence of a fair process.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowChecklist 1: the arithmetic on each document
This is the check AI does best, and where its output is easiest to verify.
- Payslip adds up. Gross pay minus deductions equals net pay. Year-to-date figures grow month by month by roughly the monthly figure.
- Deductions look plausible. Tax and contribution lines move with pay in a consistent way across months. A deduction that is a round number every month in a role with variable pay is a question.
- Bank statement balances run. Opening balance plus credits minus debits equals closing balance, and each row's running balance follows from the one above.
- Dates are possible. No payments on dates that did not exist, no weekend payroll credit from an employer that only pays on working days, no statement period that ends before a transaction it contains.
Sample prompt (use redacted text or an image with the identifiers blacked out):
Below are three monthly payslips for Applicant 1 (identifiers
removed). For each, check whether gross minus the listed deductions
equals net, and whether year-to-date totals are consistent from one
month to the next. Show your arithmetic line by line. List only
inconsistencies. Do not say whether the documents are genuine.
A sample output (illustrative) might say: "Payslip 2: gross 2,950, deductions listed total 640, net shown 2,380; calculated net 2,310, difference 70. Payslip 3: year-to-date gross is 8,850, which equals three months at 2,950, but payslip 1 shows year-to-date gross of 3,050." Two flags, both specific. You then check the arithmetic yourself, since models do slip on sums, and the flags go to the verification call. Note the instruction "do not say whether the documents are genuine": it keeps the model describing differences and not issuing verdicts.
Checklist 2: do the documents agree with each other
Fraud is rarely one forged page. It is a pack where each document looks fine and the whole does not hang together. Compare across documents:
- Employer name and spelling. The same everywhere: payslip, employer reference, contract, bank credit description.
- Pay amount against bank credits. The net pay on the payslip should appear as a credit on the statement, from a payer name that matches the employer, on about the right date each month.
- Job title and start date. Consistent between the reference, the contract and the application form.
- Address history. Addresses on the statement, the identity document and the application line up in time, and nothing moves in a way that conflicts with the tenancy history.
- Contact details on the reference. A mobile number or free webmail address for a company's HR team is a question, not a fact.
This is the most valuable use of a language model in referencing, because the volume of small comparisons is where people tire. Ask it for a table with one row per fact and one column per document, with a final column marked "agree", "differs" or "not shown".
A worked example: one applicant, three flags
An illustration with invented figures. An applicant for a flat at 1,100 a month supplies three payslips, three statements and an employer reference. The comparison table returns this (illustrative): net pay on payslips is 2,380 each month; credits on the statement are 2,380, 2,380 and 2,150, the last from a payer name that differs by one word from the employer on the reference; the reference gives a start date a year later than the first payslip's year-to-date implies.
The agent sees three questions rather than a verdict. The payer-name difference might be a payroll bureau; the lower credit might be unpaid leave; the start date could be a typing slip. The next step is a call to the employer's main switchboard (found independently, not from the reference) to confirm the applicant's job title, start date and basic pay. In this case the employer confirms all three; the payer is a payroll provider. Referencing proceeds. The point of the process is not that a fraud was found but that three possible problems were settled in a ten-minute call instead of ignored or turned into an unfair rejection.
Checklist 3: metadata and file red flags
Most PDFs carry properties such as creation and modification dates and the application that produced the file, and a PDF reader's document properties window usually also lists the fonts embedded in it. Image files carry similar data. These can be worth a look, but with caution, because a legitimate document can look odd and a skilled forger can wipe the trace.
| Signal | Possible innocent reason | What to do |
|---|---|---|
| PDF created by a desktop editor or an image editor, not by a payroll or banking system | Applicant merged files, scanned, or cropped to hide an account number | Ask for the original download or a fresh one from the portal |
| Modification date later than creation date by days | Opened and re-saved by an app | Note it; look at the content flags |
| Different fonts or spacing within one line or column | Rare in genuine payslips; can happen after conversions | Zoom in on the page and verify the amount by call |
| Same creator and timestamp on documents that claim different issuers | Applicant bundled them in a single scan or print | Ask for each document separately from its source |
| Screenshot or photo of a screen instead of a download | Applicant only has the banking app | Ask for the official statement export |
Don't ask a chatbot to read raw metadata and give a verdict. Read the properties yourself in the PDF viewer (document properties), paste the relevant text fields into the chat, and ask it to list oddities against your table. A plausible-sounding claim that "this file shows signs of editing" with no listed evidence is something to ignore. If your agency uses a dedicated referencing or document-verification service, that is a better home for this check than a general assistant, provided you understand what it examines and how it reports.
Checklist 4: patterns across a batch
If your agency handles many applications, ask whether some details repeat in ways that do not make sense. Say a three-negotiator agency receives five applications in one week for flats it manages in two neighbouring blocks: the same phone number on two employer references, the same layout flaw on three payslips from different named employers, and two applicants whose bank statements share an unusual transaction description. Individually each is fine. Together they justify escalating to a senior person, who may decide to verify all five by phone. A language model can help find these repeated details if you feed it a redacted table of the fields, but the decision to escalate is a person's.
The landlord-side checklist for AI-faked statements goes deeper on what faked statements tend to look like and how to ask applicants for better evidence.
The verification call: the step that settles it
Everything above ends here. A call to an independently sourced number settles most flags in minutes. A good call script is short:
Hello, I'm calling from [agency] about a tenancy application.
The applicant has named your company as employer. Can I confirm
three things: that they work for you, their job title, and their
start date? I'm not asking for salary details unless the applicant
has given written permission for them. Can I call you back on a
number from your main switchboard to confirm who you are?
Rules for the call: find the number yourself (company website or registered switchboard), never use a number printed on the reference; log the person's name and role, and the date and time; ask only what the applicant has consented to; and treat a refusal to confirm as information, not proof. If the contact cannot be reached, ask the applicant for another route, such as a letter on company paper sent to a verified address, or a direct employer portal confirmation if one exists.
What an honest applicant looks like when you ask
Most flagged applicants are genuine, so the way you ask matters. A message that works (illustrative):
Hi [first name], thanks for your application. One of the documents
we received didn't match another on a small detail (the payer name
on your bank credit differs from your employer's name). That's often
a payroll company. Could you tell us who runs your payroll, or send
the latest payslip direct from your employer's portal? We'll keep
your application moving in the meantime.
This version states a specific, neutral difference, offers an innocent explanation and an easy fix, and does not accuse. An applicant with a good reason answers quickly. One without often goes quiet, which is information you did not have to act unfairly to get.
Identity, landlord and guarantor documents
Payslips get the attention, but three other parts of a pack fail just as often. Apply the same flag-and-call rule to each.
Identity documents
Compare the name, date of birth and photo across the application form, the identity document and any identity or eligibility check you are required to do. A language model is poor at judging whether an identity document is genuine and should not be asked; use the checking method your agency is obliged to use, and use AI only to confirm that the typed details on the form match the details you read off the document. A transposed day and month on a date of birth is the usual honest slip, so the question to the applicant is "can you confirm your date of birth?".
Previous landlord references
A reference from a "previous landlord" is the easiest document to fake because the writer controls the letterhead. Ask the AI to compare the tenancy dates, rent figure and address on the reference with the rest of the pack: rent that does not match the applicant's stated rent, a tenancy that overlaps with another address, or an address that does not exist in the form given. Then verify by calling a number you found independently, or by asking for a tenancy agreement and rent payment evidence. For first-time renters, no previous landlord is not a flag; use a different route such as a guarantor.
Guarantors
Guarantors are referenced like applicants, and fraud is often one family member vouching for another with a borrowed payslip. Compare the guarantor's address, employer and pay with their own documents and run the same arithmetic. Check the relationship stated on the form agrees with the surnames and addresses given, and treat a guarantor who cannot be reached for a call as an open item.
A sign-off record you can copy
The record is the control that makes the rest trustworthy. Keep it to one page per application. A filled-in example (illustrative, identifiers omitted):
Application: 14B, applicant 1 and guarantor
Documents checked: 3 payslips, 3 statements, employer reference,
landlord reference, guarantor payslips
Tool and date: [assistant name and plan], checked 12 Nov
Flags raised: (1) payer name differs by one word from employer;
(2) reference start date later than payslip YTD implies;
(3) none on guarantor
Action: called employer switchboard (number from company site),
spoke to [HR contact]; confirmed title, start date, basic pay
Outcome: flags 1 and 2 explained (payroll provider; typing slip)
Decision: proceed to offer. Signed off by [senior name], 12 Nov
Notice that the decision names a person, records the evidence and shows what happened to each flag. If an applicant later asks why their application was slow, or an adviser asks how you handle automated checks, this is what you show.
Setting the threshold for when to call
Not every flag deserves the same effort, and a rule keeps the team consistent. A workable scheme for a small agency: one minor flag with a plausible innocent explanation (a payer name that differs slightly) gets an email to the applicant; two flags of any kind, or any arithmetic flag on a payslip, get a verification call; three flags, or any flag on a document the applicant cannot replace, go to the senior person before anything else happens. Review the thresholds after a month. If every application triggers a call, the checks are too sensitive and staff will stop trusting them; if none does, either your applicants are unusually tidy or the prompts are too forgiving, and a test pack with one planted error will tell you which.
What this costs in time
A quick sum, with invented figures. Say an agency in a busy month references 40 applicants, and each pack takes a negotiator about 25 minutes to read for consistency by hand. That is roughly 17 hours. With a redaction step (3 minutes), an AI comparison table (3 minutes to run) and a human read of the flags and spot-check of the arithmetic (8 minutes), the pack takes about 14 minutes: around 9 hours in total. Calls add time, perhaps 10 minutes for each of the 12 packs that carry a flag, which is 2 hours. That leaves about 5 hours a month on these figures, with the larger gain being that the consistency checking is done the same way every time. If you run the numbers on your own agency and the saving is under an hour, the process is not worth the data risk and you should stay manual.
A short script for staff: what to say if an applicant asks about AI
Applicants increasingly ask whether their documents were checked by a machine. A clear, honest answer helps. For example: "We use software to compare the documents you send with one another, for instance to check that the numbers add up. It only flags differences. A member of staff looks at every flag and may contact you or your employer. No decision is made by software alone." Make sure it is true, and that your privacy notice says the same.
How the checks fail, and how a reviewer spots it
- False flags from poor scans. A blurred "8" reads as "3" and the sums fail. Compare the flag against the page itself before calling anyone.
- Model arithmetic errors. Recheck any difference with a calculator. If the model is wrong twice in one batch, stop using it for sums and use a spreadsheet.
- Invented evidence. A model may say "font mismatch on line 12" about a file it cannot see in detail. Ask it to quote the text it is referring to, or look yourself.
- Bias creeping in. If flags cluster on applicants with irregular income, such as freelancers, or on names or documents from particular backgrounds, stop and review. Ask your adviser about the rules that apply to you.
- Everyone relaxes. After fifty clean batches, the call gets skipped. Set the rule that any flag gets a call, and sample some unflagged applications too.
For the surrounding workflow, screening tenant applications fairly with AI sets out the fairness side, and the tutorial on AI tools for landlords covers lease checks and application fraud. If you would rather map this to your agency's own referencing process, that is what my AI implementation consultation covers, on a 1:1 call.
Questions agents ask about AI and referencing fraud
Can AI tell me for certain that a payslip is fake?
No. It can flag inconsistencies: arithmetic that does not add up, fonts that change mid-line, dates that clash, a pay pattern that does not match the bank credits. Those are reasons to verify, not proof. Only the employer, bank or tenant can settle it.
Is it fair to reject an applicant on an AI flag?
Don't. Treat a flag as a question and ask the applicant for an explanation or an alternative document. An automatic rejection risks refusing honest applicants whose documents look odd for ordinary reasons, and it may fall foul of equality and data rules. Ask your adviser what applies to you.
Do metadata checks prove tampering?
No. File properties show which software created or last changed a PDF, and when. An edit tool in that field is a prompt to ask questions, since many genuine documents are scanned, merged or exported by apps. Combine it with content checks and a call.
Should the tenant be told AI is used?
Be open about it. Say in your referencing notice that documents may be checked using automated tools and that a person reviews any concern. Your privacy notice should describe it accurately, and an adviser can confirm the wording for your business.
Further reads
- Renters' Rights Act: Using AI for Tenant FAQs and Section 13 Letters — Tenant FAQs and formal letters with AI.
- How Can AI Help in Property Management? Tenant Messages and Repairs — Tenant messages and repairs once a tenancy starts.
- AI Assistant for Landlords: WhatsApp Maintenance Requests and FAQs — Maintenance requests and FAQs on WhatsApp.
- AI for Letting Agents: Handle Tenant Queries Without Extra Staff — Handling tenant queries without extra staff.
- AI for Rental Properties: Screening Enquiries, Viewings and Tenant Replies — Screening enquiries and viewings before referencing starts.
- Tenant Complaint Response Letters When ChatGPT Wrote the Complaint — When a complaint letter was written by ChatGPT.
- AI Loan Processing: Catch Missing Documents Before Underwriting — A small-brokerage routine: classify each file's documents, test them against a loan-type checklist, run stale-date formulas and have a processor sign off.
- Landlord Compliance Checklist: Track Gas Safety and EICRs With AI — Read certificates with AI, log expiry dates in one tracker, set reminders and keep a short exception list, with a human check on every certificate.
- Letting Agent Software With AI: A Checklist Before You Switch — A vendor question list for letting agencies: data export, client money, what the AI really does, per-property pricing and a demo script to score.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: general document-forensics practice (file properties, arithmetic and consistency checks); AI provider data-use pages; the author's checklist design. Sample documents and figures are invented and illustrative. Check equality, data-protection and referencing rules with a qualified adviser.