Ask the agent for inconsistencies, not accusations. It tests payslips and statements for internal arithmetic and tampering, compares the application with bureau and account data, checks device and velocity signals, and confirms who owns the disbursement account. It writes up what it found. A fraud analyst decides whether there is fraud, what marker to record and whether the application stops.
What is application fraud review, and who owns it?
Application fraud is lying to get credit. It comes in a few shapes. First-party fraud is a real applicant inflating income or hiding debts. Identity fraud is someone applying in another person's name. Synthetic identity fraud combines real and invented details into a new person. Stacking is one applicant taking loans from several lenders at once before any of them sees the others. Each leaves different traces.
The fraud team owns the review, usually working alongside lending operations. Applications are referred by rules, scores or underwriters. The agent's value is in the slow part: reading documents line by line and comparing every data point, which analysts rarely have time to do on every referral.
What does the agent prepare?
- Agent, depth 1: tests documents. Payslips: does gross minus deductions equal net, are tax and other deductions plausible for the stated pay, do dates and pay periods line up. Bank statements: does the running balance add up line by line, do salary credits match the payslips.
- Agent, depth 1: checks file properties. Document metadata, edits after creation, fonts or layouts that differ within one document.
- Agent, depth 1: compares sources. Application data against bureau data, existing account data and earlier applications. Address history, employer, income and declared debts.
- Agent, depth 1: reads device and velocity signals. Devices or contact details seen on other applications, many applications in a short time.
- Agent, depth 1: checks the disbursement account. That it is held in the applicant's name. A disbursement account in another name is one of the clearest warning signs.
- Agent, depth 2: drafts the fraud summary. Each finding with its evidence, a list of innocent explanations considered, and the questions an analyst might ask the applicant or employer.
- Person: decides. The analyst decides whether there is fraud, whether to verify with the employer or the applicant, whether to stop the application and whether to record a fraud marker.
- Person: handles identity theft victims. If someone else's identity was used, contact with the real person is handled with care by a person.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Documents | Tests arithmetic and consistency | 1 | Judging a finding |
| File properties | Checks metadata and edits | 1 | None |
| Compare | Cross-checks all sources | 1 | Explaining differences |
| Signals | Reads device and velocity | 1 | None |
| Disbursement | Confirms account ownership | 1 | Accepting a mismatch |
| Summary | Drafts findings and questions | 2 | The fraud decision |
| Outcome | Nothing | None | Stopping, declining, markers |
| Victims | Nothing | None | Contacting the real person |
Why do the decisions stay with an analyst?
A fraud finding has lasting consequences. A marker recorded against someone can follow them across lenders. If the agent is wrong, an honest applicant may find credit closed to them for years without knowing why. That calls for a named person who weighs the evidence and owns the call.
Stopping an application is also a decline in effect, and declines are a credit decision. Whether an agent can decline a loan application is answered no for that reason. And where the applicant is a victim of identity theft, the case turns into supporting that person, which is no job for an agent.
Which controls apply, and what evidence do they leave?
Findings, not verdicts. The agent's summary describes what does not add up. It does not label the applicant a fraudster. Evidence: the summary text, checked at certification under the scope gate.
Innocent explanations listed. Small employers produce unusual payslips. Bonuses make salary credits jump. The agent records which innocent explanations it considered and why each does or does not fit.
Provenance. Documents uploaded by the applicant are declared. Bureau data is external. Your own account data is sealed. A conflict between declared and sealed data is a strong finding; a conflict between two external sources is weaker.
No customer contact. The agent does not tell the applicant its findings. Revealing which check failed teaches a fraudster what to fix next time.
Analyst decision filed. Outcome, marker and reasons are filed with the analyst's name.
What should you measure?
- Referrals confirmed as fraud, by the finding that triggered them.
- Fraud found later on applications the agent reviewed and passed. Each one is a certification case.
- Honest applicants stopped. Track appeals and markers removed.
- Time from referral to decision.
- Findings rejected by analysts, by type. Shows which tests are noisy.
Where does it usually go wrong?
Formatting mistaken for forgery. A payslip from a small business looks unusual because it was made in a spreadsheet. Arithmetic errors matter more than layout.
Stacking missed. Bureau data lags. An applicant with several pending applications elsewhere may look clean. Velocity signals and recent searches help; the agent should state the date of the bureau data it used.
Over-reliance on one signal. A shared device can be a family computer. The summary should rank findings by strength and say which stand on their own.
Card and application fraud blurred. A stolen card on an existing account is a different case, handled as in reviewing card fraud alerts.
The non-fraud path for the same applications is in processing personal loan applications. The general bounds are in AI agents in banking.
Questions people ask
Can the agent verify employment by calling the employer?
It can draft the verification request and find contact details from independent sources, not from the application. A person makes the call or sends the request.
What if the documents look fine but the bureau shows a different address?
The agent records the difference with both sources and their dates. People move, and bureau data lags; the analyst decides whether it matters.
Should the agent review every application or only referrals?
Document arithmetic checks are cheap enough to run on every application. The full review, with a summary for an analyst, is for referrals.