How AI is changing financial services analytics

Financial firms were early users of statistical models, long before anyone called them AI. What changed in the last few years is the range of data those models can use, the tasks they can take on (reading documents, triaging alerts, summarizing cases) and the number of smaller firms that can now afford them. This post looks at four areas where AI is used in financial services analytics, the rules that shape how it can be used, and practical first steps for a credit union, lender, fintech or advisory firm without a large data science team.

1. Fraud detection

Fraud detection is the most established use. The task: find the tiny share of fraudulent transactions among a very large number of legitimate ones, fast enough not to slow down payments.

Models work from features built out of raw transactions:

• Velocity: how many transactions this card or account made in the last hour, day or week.
• Geography: distance from the previous transaction, whether the location fits the customer's pattern.
• Behavior: how far the amount, merchant type or time of day is from the customer's normal.
• Network: shared devices, addresses or payees linking accounts that should not be linked.

Common model families are gradient-boosted trees for scoring and isolation forests or other anomaly detectors for unusual patterns. Many systems score every transaction with a fast model and pass only the suspicious ones to slower, more thorough checks.

Why false positives matter more than detection rates

A small worked example shows the real problem. Say a processor sees 1,000,000 transactions a month and 0.1% are fraudulent: 1,000 frauds.

• A model catches 90% of them: 900 frauds flagged.
• It also flags 1% of legitimate transactions: 1% of 999,000 = 9,990 false alarms.
• Of the 10,890 alerts, only 900 are fraud. Precision is about 8%. Investigators spend over 90% of their time on legitimate customers.

Cut the false positive rate to 0.5% and false alarms fall to about 5,000, so precision nearly doubles to about 15% with the same detection rate. That is why fraud teams judge models on alerts per confirmed case and on customer friction, not only on "percentage of fraud caught". When a vendor quotes a detection rate, ask for the false positive rate on data like yours.

2. Credit risk

Traditional scorecards use a small number of well-understood factors from credit bureau data: payment history, amounts owed, length of credit history, new credit and credit mix. They are transparent and well tested, but they cannot score people with thin or no credit files. The CFPB has estimated that tens of millions of U.S. adults fall into that group.

Machine learning models can use many more variables, including cash-flow data from bank accounts (income regularity, balances, how bills are paid), rent and utility payments. Lenders use this to assess applicants a scorecard would reject for lack of history.

The risk is that more variables make it easier for a model to pick up proxies for protected characteristics. A postcode or a shopping pattern can stand in for race or age even when those are excluded. That is why fairness testing (comparing approval rates and error rates across groups) and documentation of which features were used, and why, are now standard practice rather than extras.

3. Compliance and anti-money laundering

Rule-based transaction monitoring flags anything over fixed thresholds and produces large volumes of alerts, most of which turn out to be benign. AI is used in three ways here:

• Alert triage: scoring alerts so investigators see the likely suspicious ones first.
• Network analysis: finding structures that rules miss, such as many small deposits spread across linked accounts to stay under reporting thresholds.
• Document work: extracting fields from contracts, onboarding documents and filings, and flagging new regulatory publications that may apply to the firm.

The human decision stays: a person still decides whether to file a suspicious activity report. AI changes where their time goes.

4. Customer analysis

• Churn prediction: falling balances, fewer transactions and more complaints often come before a customer leaves.
• Next product: transaction patterns can suggest a customer is saving for a house or starting a business.
• Complaints and feedback: language models can group complaint text by theme so recurring problems show up in numbers rather than anecdotes.
• Pricing: rates and fees adjusted by risk and competition, within what regulation and fairness allow.

The rules that shape how AI can be used

This is a summary, not legal advice; check the current rules for your jurisdiction and charter.

• Explanations for adverse decisions. In the U.S., the Equal Credit Opportunity Act and Regulation B require lenders to give applicants the specific principal reasons for a denial or worse terms. The CFPB has said these requirements apply in full when the decision comes from a complex model. "The model scored you as high risk" is not a reason; "high credit utilization" is.
• Model risk management. The Federal Reserve and OCC guidance known as SR 11-7 sets expectations for developing, independently validating, monitoring and documenting models. Examiners apply it to machine learning models as well.
• The EU AI Act classifies AI systems used to assess the creditworthiness of individuals or set their credit scores as high-risk (systems used only to detect financial fraud are excluded from that category). High-risk systems carry obligations for risk management, data quality, documentation, human oversight and accuracy, which are being phased in.
• Fair lending. Disparate impact testing before deployment and on a schedule afterwards, with records of what was tested and what changed.

In practice this pushes many firms toward models they can explain: gradient-boosted trees with feature-level explanations such as SHAP values, or carefully built logistic regression, rather than the most complex model available.

What gets in the way

Data spread across old systems. Customer data in one system, transactions in another, products in a third, often with different IDs and formats. Most AI projects that disappoint fail on data, not on the model. Clean and connect the data first.

Explainability and validation costs. Every model that affects customers needs documentation, validation and monitoring. Budget for that work, not only for building the model.

Skills. People who understand both the models and the regulations are scarce. Training staff who already know the products and rules is often more effective than hiring pure technologists.

Where a smaller firm should start

1. Pick one use case with a measurable result. Alert triage (hours saved per case), churn (members retained), document extraction (hours per application). Measure the baseline first.
2. Buy decisioning, build understanding. For fraud scoring and credit decisions, established vendors with explainability and validation support are usually a better route than building in-house. Keep enough internal knowledge to question their results.
3. Get the reporting right. Before any model, make sure you can answer basic questions reliably: volumes, balances, losses, complaints, by month and by product. Models built on numbers nobody trusts will not be trusted either.
4. Write down how you will govern it before the first model goes live: who approves it, who validates it, what triggers a review.

Where clariBI fits, and where it does not

clariBI is a reporting and analysis tool. It is not a fraud scoring system, a credit decisioning engine or a compliance platform, and it holds no certifications such as SOC 2. It has no connection to core banking, trading or custody systems, and you should not load customer financial records you are not allowed to share with a processor.

What it does help with is step 3 above and the analysis around a firm's own business:

• Market and economic data. API connectors for FRED, Twelve Data, Alpha Vantage, Financial Modeling Prep, the U.S. Bureau of Labor Statistics and others, so rates, prices and macro series can sit next to your own numbers. The investment research use case lists them.
• Business performance. Stripe, PayPal, Square, Ramp and Plaid through the MCP catalog, plus spreadsheets and exports for everything else; the finance teams use case shows common questions.
• Forecasts with ranges. On Starter and above, forecasting uses statistical methods chosen by backtesting, with 50%, 80% and 95% bands and flags for anomalies and changepoints. No language model is involved in the calculation. The post on anomaly detection for business metrics explains how those flags work.
• Controls. Roles and two-factor sign-in on every plan, and an organization audit trail on Professional and Enterprise.

When you use the AI features, the language model receives column names, a few sample values, your question and the calculated results, not your full dataset. The help article on when AI analysis falls short covers how to check answers. For a closer look at risk scoring and compliance reporting specifically, see financial services analytics: risk, compliance and customer intelligence.

Next step

List the five numbers your leadership asks for every month (for example loan volume, delinquency rate, deposits, complaints and cost per account). For each, write where it comes from today and how long it takes to produce. The slowest one is where analytics, AI or not, will pay off first.