Let's Talk
AI Fraud Detective — app mockup

🧪 Experiment No. 01

AI Fraud
Detective

A gamified fraud detection simulator exploring how AI can guide user decisions in fintech — combining decision-making under uncertainty, pattern recognition, and progressive learning.

Type AI Product Experiment
Domain Fintech · Gamification
Stack Lovable · Prompt Engineering
Status v4 — Completed

Problem

Fraud detection systems are invisible and complex. Users don't understand why transactions are flagged — creating low trust, poor awareness, and reactive behavior.

Hypothesis

If fraud detection becomes interactive, explainable, and slightly gamified, users can better understand risk patterns and make more informed decisions.

Fraud systems are invisible to the people they're meant to protect

Most fraud detection happens behind the scenes. When a transaction gets flagged, users receive a binary alert — blocked or approved — with no explanation. They don't understand what triggered the flag, what patterns look suspicious, or how to protect themselves proactively.

This creates a compounding problem: low trust in financial systems, poor user awareness of actual risk signals, and behavior that stays reactive rather than informed.

"What if the fraud detection interface was designed for the user, not just the system?"

A simulator that teaches through doing

I designed a gamified fraud detection simulator where users step into the role of a fraud analyst. Instead of passive alerts, they actively review transactions, classify threats, and receive AI-style explanations for each decision.

The experience progressively escalates in complexity — from basic pattern recognition to multi-signal investigation — mirroring how real fraud analysts develop their instincts over time.

Transaction review AI-generated explanations Progressive difficulty Pattern recognition Connected scenarios Score & feedback loops

Four systems working together

🧾

Transaction Analysis

Users evaluate each transaction across multiple dimensions — amount, location, device fingerprint, and behavioral signals. Each case is unique, avoiding repetition.

⚠️

Risk Signals

Transactions surface contextual signals: unusual location, new device, high amount, abnormal timing. These guide decision-making without giving away the answer.

🧠

AI Guidance Layer

The system provides proactive hints, post-decision explanations, confidence levels, and pattern recognition feedback — simulating how a real AI co-pilot would behave.

🔗

Connected Scenarios

Fraud is presented as linked events — a suspicious login followed by a test transaction followed by an attempt. Users experience fraud as a chain, not isolated incidents.

🎮

Gamification Layer

Score, streak, levels, and badges create engagement without undermining seriousness. The gamification is subtle — it rewards understanding, not just speed.

📈

Progressive Learning

The experience evolves from simple binary decisions to complex multi-signal investigation, mirroring real skill development.

See it in action

Mission Briefing

Onboarding flow — how the user enters the experience and receives their objective.

Transaction Review

Core loop — evaluating a transaction, reading risk signals, and making a decision.

AI Feedback & Score

Post-decision AI explanation, confidence breakdown, and score update.

Four versions to find the right balance

Each version surfaced a different failure mode in the design thinking.

v1

Simple Classification

Basic transaction approve/reject flow. Too linear and repetitive — users ran out of engagement before they ran out of transactions.

v2

Gamification Added

Introduced score, streak, and case variation. Improved engagement but the experience still felt like a quiz rather than a simulation.

v3

Dynamic Events & Levels

Added progressive difficulty and dynamic event generation. Users felt progression — but decisions still felt isolated from each other.

v4

Connected System (Final)

Added AI hints, memory across scenarios, connected fraud chains, and uncertainty states. The experience now felt close to a real fraud detection system.

What this experiment revealed

1

UX alone is not enough

Good interface design does not create engagement without variation and progression. The structure of the experience matters as much as how it looks.

2

Perceived intelligence matters

Users engaged significantly more when the system felt proactive, contextual, and aware of past behavior. The AI layer didn't need to be actually intelligent — it needed to feel that way.

3

Fraud is about patterns, not events

Moving from isolated decisions to connected scenarios was the single biggest improvement. Every decision felt more meaningful and realistic.

4

Gamification requires precise calibration

Too much = trivial. Too little = work. In fintech especially, the balance is critical — the subject matter demands respect even in a simulation.

Where this could go in a real product

This was a design experiment, not a production product. But the concepts have direct application in real fintech contexts:

🏦

User Fraud Awareness

Banks could embed a lightweight version inside their app — turning fraud alerts into learning moments rather than dead ends.

📉

Reduced Support Costs

Users who understand fraud signals are less likely to call support after a block. Education at the point of friction reduces load downstream.

🤝

Increased Trust

Transparency about why decisions are made builds significantly more trust than opaque automated systems.

🎓

Staff Training Tool

A more sophisticated version could onboard junior fraud analysts faster than traditional documentation-heavy training.

Real-time transaction simulation Banking API integration Adaptive AI explanations Personalisation by behavior Multi-player mode for analyst teams

Built with speed in mind

This experiment was built using a prompt-driven workflow — rapid prototyping with Lovable, iterated through UX design thinking, and evaluated against product criteria at each version.

Lovable Prompt engineering UX design thinking AI-assisted logic simulation Rapid iteration
← Back to AI Lab All work →