Fraud Detection with AI — a lender's playbook

Most fraud content is either (a) vendor marketing disguised as education, or (b) an academic taxonomy that can’t help you at 9:30am on a Monday when a ring is pulling money out the back of your platform. This playbook is neither. It’s the compact, opinionated, field-tested version of what I’d hand a new fraud lead on day one.

The four patterns

Every fraud that will materially hurt a small-to-mid-sized lender belongs to one of four families. The playbook walks each one with a sanitised real-world example, the AI prompts that detect it, the decision tree that wraps the detection, and the signature false-positives you’ll need to tolerate or filter.

1. Synthetic identity. Fake humans built from real pieces — real PAN, real phone, real address, combined in a way the bureau has never seen before. The fastest-growing pattern globally. The prompt that catches these isn’t looking at the application; it’s looking at the timing and consistency of the digital breadcrumbs.

2. Document manipulation. Payslips, bank statements, employment letters — small, surgical edits that fool a tired underwriter but leak a pattern when the full document is read in context. AI is surprisingly good here, but it has three signature false-positives you must know.

3. Ring fraud. Coordinated applications — different names, different PANs, same address or same employer or same repayment account. Easy to see in aggregate, nearly impossible to see one loan at a time. The playbook includes the specific SQL patterns and AI dossier prompt that surface rings early.

4. Re-verification and payout-stage scams. The loan is approved to a genuine borrower; then, between approval and disbursal, the bank account changes, or the OTP moves, or the phone number is redirected. This is a workflow problem more than an underwriting one, and the playbook spends a full chapter on the workflow fixes.

The prompts, and where they break

Twelve prompts. Each prompt one page: goal, prompt text, input spec, an example where the prompt caught a real (sanitised) fraud, and an example where it flagged a false positive. The “false positive” page is the one you’ll come back to most — it’s how you tune the prompt to your population without drowning in noise.

The decision trees

Three of them, printable, one page each. Onboarding-stage (does this borrower make it to underwriting?), payout-stage (does the approved loan actually disburse?), and post-disbursal (did something we should have caught slip through?). Each tree combines AI signals with deterministic rules, because the combination beats either alone.

The hardest chapter

“What AI still misses.” A short, honest chapter on the fraud patterns current LLMs can’t reliably catch: sophisticated deepfake video KYC, collusive employer fraud, and insider-assisted fraud. If I sold this playbook as “AI catches all fraud” I’d be lying. These categories need different defences and I tell you what they are, briefly.

The case study

One full teardown. A synthetic-ID ring that cost multiple Indian NBFCs real money in 2024. Six signals, across three stages, that would have caught it — if anyone had been looking for them together. The teardown is the part of the playbook people quote most.

Where this sits

If you bought the $29 Prompt Library, the fraud chapter there is ~7 prompts. This playbook is the deeper version: 12 prompts, the patterns around them, and the decision trees that make them actually operational. If you want library + scorecard + playbook + starter-pack, the $149 Starter Kit bundles all of them.