— What we hear
"Our fraud system is succeeding at fraud and quietly killing onboarding."
False-positive rates that look small at the model layer become large at the customer-journey layer. The board sees fraud losses; the support inbox sees the rest.
"We can't put this model in production until risk signs off, and risk won't sign off until we can explain every decision."
SR 11-7, EU AI Act, FCA's AI guidance — the model risk frameworks are real, and they reject black boxes by default. Most teams discover this after the model is built.
"We've got six AI initiatives across three business units and none of them share infrastructure."
FinTechs at scale develop AI capability the way they developed monoliths — fast, isolated, and unmaintainable. The platform conversation always comes a year late.
"Compliance and product can't agree on what 'AI use' even means for disclosure purposes."
The regulatory definition of "automated decisioning" is broader than most product teams realise. The conservative read is the safe read — but it requires UI changes, consent flows, and disclosures most products haven't built.
— Pattern recognition
— What fails
Binary block/allow decisions. The boundary between fraud and not-fraud is fuzzy; pretending it's binary is what produces both false-positives and slipped fraud.
Models that can't show their work. If the compliance team can't trace a decision to its features, the model doesn't deploy. Build for explainability from day one.
One-shot launches. Production deployment without a shadow-scoring phase produces unhappy surprises that compliance can't unsee.
Ignoring downstream metrics. Fraud rate is the goal. Funnel conversion, support ticket volume, and 30-day retention are the constraints. Optimise on the goal and the constraints will move against you.
— What works
Tiered confidence with escalation paths. Block, allow, step-up auth, human review. Four tiers convert most of the "false-positive" volume back into successful customers.
Shadow scoring before takeover. Run the new model alongside the old for 4–8 weeks. The risk committee gets data to override on; the engineering team gets confidence; the regulator gets a paper trail.
Feature stores with audit lineage. Every prediction needs to be reproducible 18 months later from the exact features that produced it. This is non-negotiable for SR 11-7 compliance.
Human-in-the-loop where the stakes warrant it. Credit decisions over a threshold, sanctions hits, high-value transactions. The loop is not a fallback — it's the design.
— Engagement patterns
— 01
Rebuilding decision layers that succeed at fraud without killing legitimate transactions. Tiered scoring, shadow validation, model risk documentation.
— 02
Reducing analyst load on alert review without missing structurally suspicious patterns. NLP on documents, entity resolution, network analysis.
— 03
Adverse action notice-ready credit models with full feature attribution. Alternative data integration, explainability layer, fair-lending audit.
— 04
Agentic workflows for ticket triage, refund disputes, document processing. Built with compliance-grade audit trails for every action taken.
— Series C payments platform (illustrative)
A Series C consumer payments platform was losing customers to fraud-system friction. The team had tried two rule engines and a vendor model. We rebuilt the detection layer in 14 weeks — false-positives dropped 41%, true-positive recall held flat.
— Composite case drawn from engagement patterns; anonymized for client confidentiality. Real metrics, real outcome arcs.
— If this resonated
If you're a FinTech leader sitting with one of the problems on this page, applications are open for next quarter. We read every one.