Your fraud team reviews 40 cases per day. Every case they review is a genuine decision — a transaction pattern that the model flagged because it deviates from the account's baseline behavior in a way the rules couldn't explain. They decline 34 and clear 6. The 6 clearances take 4 minutes each. The 34 declines are logged with reason codes that feed back into the model. Your false positive rate on legitimate transactions is 0.3%. Your fraud loss rate is 0.08% of GMV. Neither number was achievable when every transaction ran through a static rules engine.
I've been watching payments companies try to solve fraud with rules. A rules engine is a list of conditions that a fraud team agreed on 18 months ago. Fraudsters iterate faster than rules reviews. By the time the rules team meets to add a new pattern, that pattern has already cost 6 weeks of losses. The rules also don't know that a transaction that looks unusual for the account as a whole is completely normal for that account on Saturday evenings. Context is the thing rules don't have.
The AI Fraud Detection Maturity Ladder for Payments Companies
Stage 1: Transaction feature engineering. Every transaction is tagged with behavioral context — time since last transaction, deviation from average transaction size, merchant category frequency, device fingerprint, geolocation delta, and velocity across the last 24 hours. These features are calculated at transaction time, not batch-processed overnight. The feature store is the foundation everything above depends on.
Stage 2: Behavioral baseline per account. Each account has a behavioral model — typical transaction size, typical merchant categories, typical transaction times, typical device and location. A transaction is scored against the account's own baseline, not against a population average. A $4,000 transaction from a business account that regularly transacts at $3,000-$5,000 scores low. The same transaction from a consumer account that typically transacts at $50 scores high.
Stage 3: Real-time ML scoring. The fraud model scores every transaction at authorization time — not after the fact. The score, combined with the transaction amount and account type, determines the action: approve, decline, step up to 2FA, or route to manual review. Decisions happen in under 150ms so the payment experience for legitimate users isn't degraded.
Stage 4: Case management and feedback loop. Every manual review decision feeds back into the model — confirmed fraud, false positive, or cleared. The model retrains on new cases weekly. A fraud pattern that first appeared in manual review is captured in the model within 7 days. The fraud team's review decisions become training data, not just operational overhead.
Stage 5: Chargeback and dispute automation. Transactions that result in chargebacks are analyzed for model signal — did the model score them correctly? Chargebacks on transactions the model missed update the training data and trigger a rules review. Disputes that follow a pattern the model recognizes as friendly fraud are flagged automatically. The chargeback team's workload concentrates on genuine disputes.
What Each Stage Changes
Stage 2 is where false positives drop. A model that compares transactions to an account's own baseline declines far fewer legitimate transactions than one that compares to a population average. False positives are the hidden fraud cost — every legitimate transaction declined is revenue the merchant loses and a customer the payments company risks. Stage 4 is where the model compounds. Each week of review decisions improves accuracy. After 6 months, the model is materially better than it was on launch day. Stage 5 is where the economics become visible in P&L — fraud loss rate and chargeback rate declining simultaneously.
Wednesday's Track Record
Wednesday Solutions has built data and AI systems in production for fintech and financial services companies, and worked with engineering teams at American Express, Visa, Capital One, and Discover on payment-side infrastructure. The real-time feature engineering, behavioral baseline architecture, ML scoring pipeline, and feedback loop required for a payments fraud detection system is work the Wednesday team has delivered in environments where latency and accuracy are both hard constraints.
Sachin Gaikwad, Founder & CEO at Buildd: "Wednesday Solutions' team is very methodical in their approach. They score very well in terms of the scalability, stability, and security of what they build."
The Entry Engagement
The Wednesday team starts with a 2-week fixed-price evaluation sprint. They analyze your current transaction data, map your existing rules engine logic, and deliver a false positive rate and fraud detection rate benchmark for the current system versus a baseline ML model. If the model doesn't show a clear path to 50% cost reduction — through false positive reduction, fraud loss reduction, or review team efficiency — the evaluation stops and you don't pay for the build.
Talk to the Wednesday team — Send them your current fraud loss rate, your false positive rate, and your manual review volume. They'll tell you what a model can and can't improve before you commit.
Top comments (0)