AI underwriting and fraud detection in lending

The fuel is alternative data: signals beyond the credit bureau. Common examples:

• Cash flow data from a linked bank account (via an aggregator like Plaid): income regularity, overdraft frequency, savings buffer.
• Rent and utility payments, which predict repayment but rarely appear in FICO.
• Employment and payroll verification.

Cash flow underwriting is now a recognized practice. The CFPB has publicly discussed cash flow data as a way to expand access to credit, and open banking rules finalized in the US are pushing consumer-permissioned data into the mainstream.

What the model actually outputs

The model does not output "approve" or "deny." It outputs a probability of default (PD): the estimated chance the borrower fails to repay over a set window. The lender then combines PD with loan amount, term, and its own cost of capital to price the loan (risk-based pricing).

The model estimates risk. Humans and policy set the thresholds and prices.

The regulatory catch: you must explain the decision

Here is where fintech differs from most AI applications. Under the US Equal Credit Opportunity Act (ECOA) and its rule Regulation B, a lender that denies credit must give the applicant specific reasons. "The algorithm said no" is illegal.

This creates a hard constraint: the model must be explainable. Teams use tools like SHAP (SHapley Additive exPlanations), which attributes a prediction to individual features, to generate the required adverse action notices ("declined due to: high recent credit utilization, short employment history").

Two more constraints matter:

• Fair lending / disparate impact. A model may not use race, gender, or a close proxy. Even a neutral variable (say, zip code) can be an illegal proxy if it produces a disparate impact on a protected group. Lenders run fairness testing before deployment.
• Model risk management. Regulators expect documented validation, monitoring, and governance, echoing the Federal Reserve guidance known as SR 11-7.

The lesson: in lending, a slightly less accurate model that you can explain and defend often beats a black box you cannot.

🎬 [VIDEO: "Explainable AI with SHAP" — youtube.com — a clear walkthrough of how SHAP attributes a model's prediction to each input feature]

Fraud detection: a different problem entirely

Underwriting asks "will this person repay?" Fraud detection asks "is this person even real, and is this transaction legitimate?" The two run at different speeds. Underwriting can take a minute. Fraud scoring must happen in milliseconds, before money moves.

Two flavors of fraud

• First-party fraud: a real person lies (inflates income, never intends to repay).
• Third-party fraud: someone uses a stolen or synthetic identity. Synthetic identity fraud, stitching a real Social Security number to a fake name and history, is one of the fastest-growing types and is notoriously hard to catch because the "person" looks new, not fraudulent.

Transaction graphs and catching fraud rings

Single-transaction rules ("flag any transfer over $5,000") catch clumsy fraud. Organized fraud rings evade them by spreading activity across many small accounts. To catch rings, you have to look at relationships, not just transactions.

Enter the transaction graph. Model every account as a node and every transfer, shared device, shared phone number, or shared bank as an edge. Fraud rings light up as dense clusters: fifty "different" applicants sharing three devices and funneling funds to one account.

Graph neural networks (GNNs) learn patterns over these structures. Instead of scoring an account in isolation, a GNN incorporates the behavior of its neighbors. An account that looks clean alone but sits one hop from ten known mule accounts gets flagged.

A simple pattern lenders watch for:

• Many new accounts, created within a short window, from overlapping device fingerprints.
• Funds that fan in to a small number of cash-out points.
• Application data that is subtly recycled (same employer, same address formatting, sequential emails).

Speed matters: real-time scoring

Fraud decisions happen inline. A typical flow:

1. Application or transaction arrives.

2. Features are pulled from a feature storefeature storeA centralised repository managing ML features, ensuring consistency between training and serving environments.View full definition → (precomputed signals kept fresh, like "device seen on 8 accounts this week").

3. A model returns a risk score in milliseconds.

4. Low score: proceed. Medium: step-up authentication (extra verification). High: block and review.

The graph is precomputed and updated continuously so the live decision stays fast. The tradeoff every fraud team manages: block too aggressively and you create false positives (blocking real customers, called "friction"), block too little and losses rise.

Putting it together: the digital lender's stack

Trace one applicant through a modern digital lender:

1. Identity and fraud check first. Is this a real person? Does the device, email, and application pattern resemble a known ring? A GNN plus rules score this in real time. Fail here and underwriting never runs.

2. Data aggregation. With the applicant's permission, pull bank transaction history and verify income and employment.

3. Underwriting model. Gradient-boosted model outputs a probability of default from bureau data plus alternative data.

4. Pricing and policy. Convert PD into an approve/decline and a rate, respecting fair lending rules.

5. Explainability layer. Generate the adverse action reasons if declined, using SHAP or similar.

6. Ongoing monitoring. Watch for model drift (performance decay as the economy or borrower mix shifts) and for fraud rings adapting to the model.

Notice that AI appears at two very different points, doing two very different jobs, under two very different time budgets.

Where teams get it wrong

• Chasing accuracy over explainability. In lending, an unexplainable model is a compliance liability, not an asset.
• Ignoring drift. A model trained on a low-rate, low-unemployment period can quietly degrade when conditions change. Monitoring is not optional.
• Treating fraud as static. Fraudsters adapt within weeks. A model that is not retrained becomes a mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → of last quarter's attacks.
• Data leakage in fraud graphs. Using information that would not actually be available at decision time inflates test performance and collapses in production.

Key takeaways

• AI underwriting expands access by reading alternative data (cash flow, rent, payroll) that FICO ignores, turning "thin file" applicants into scorable borrowers.
• Gradient-boosted models output a probability of default, not a yes or no. Business rules and risk-based pricing turn that probability into a decision.
• Explainability is a legal requirement, not a nicety. ECOA and Regulation B force lenders to justify every denial, so tools like SHAP and fair lending testing are core to the stack.
• Fraud detection runs in milliseconds and depends on relationships. Transaction graphs and graph neural networks catch fraud rings that single-transaction rules miss.
• Both systems decay without monitoring. Model drift and adapting fraudsters mean retraining and governance are ongoing, not one-time, work.