Fintech Security & UI Design

Tomaan: Fraud Verification

Exploring how probabilistic AI risk assessment could be layered onto an offline OTP authentication flow.

Role Product Designer
Status 2026 Concept Exploration
Tomaan Security Dashboard Showcase

Overview

This concept asks a narrow, difficult question: can a real-time AI fraud check sit on top of an offline one-time-password (OTP) system without weakening the one guarantee that makes it trustworthy? That guarantee is that your code never has to travel anywhere to prove who you are.

Context: The Legacy Offline OTP Foundation

At Carrene, I designed the interfaces for offline OTP apps used to protect card transactions for a security-conscious banking user base. Those apps scaled past 2 million users. The core mechanism generated a code entirely offline, valid for 60 seconds. Tomaan explores adding a generative AI judgment layer to this proven system to stop authorized fraud without ever touching the offline guarantee underneath it.

The UX Challenges of AI Security

Three things are true about generative security checks that were not true about deterministic offline flows:

Challenge 1

Generative Latency

They are slow relative to user expectation. AI checks take time, and a static loading spinner over that duration reads as "broken," which destroys systemic trust.

Challenge 2

Estimate Uncertainty

They are probabilistic. The old flow did not guess, a code either matched or it did not. An AI fraud check only estimates, and that uncertainty must be shown clearly.

Additional Challenge: Quiet Failures

An AI layer can fail quietly by timing out, degrading, or hallucinating. The interface needs a way to say "this part of the system isn't sure" without triggering panic.

Design Decisions & Trade-offs

I designed three core screens that answer these constraints directly, focusing on human control and recovery states.

Tomaan Dashboard Screen
01: Dashboard

Visual Trust

Constraint: AI estimates bring uncertainty to a deterministic system.
Alternatives: I considered full-screen alerts, hiding the AI assessment entirely, or using an ambient colored chip.
Decision: I chose the ambient chip under the balance.
Trade-off: Users might ignore it over time, but it acts as a low-stakes exposure to AI certainty without blocking their main task.
Tomaan Authorization Screen
02: Authorization

Designing for the Wait

Constraint: Generative latency makes users think the app is broken.
Alternatives: I considered a standard loading spinner, a categorical risk label, and a precise numeric confidence score.
Decision: I chose a typewriter-style sequence that narrates the check, resolving into a categorical label ("Low Risk"). I actively avoided numeric scores like 98% because they imply a false mathematical certainty that this hypothetical model cannot guarantee.
Trade-off: The text takes up more visual space, but it keeps the user informed and sets honest expectations about the AI's capabilities.
View Fallback State (Handling Quiet Failures)
Tomaan Fallback Screen

03: Precision Degradation

Constraint: An AI layer can fail quietly or time out.
Alternatives: I considered blocking the transaction entirely, showing a generic error message, or falling back silently.
Decision: I chose to explicitly name the failure ("Recipient check unavailable") and default back to the original offline OTP mechanism.
Trade-off: It breaks the seamless AI experience, but a fallback to a proven legacy mechanism is not an error. It is the system working correctly as a safety net.

Takeaways & Open Questions

If bringing this into a real production environment, I would resolve these UX risks with the security and engineering teams:

Validation 1

Alert Fatigue

Confidence indicators need to earn their visibility. If every transaction shows a badge, users stop reading them. We need usage data to decide how often to surface the chip versus staying silent.

Validation 2

The Missing Middle

This designs the happy path and the timeout path. The missing state is genuine uncertainty. My proposal is a distinct screen requiring a slow-press hold confirmation that names specific doubts in plain language.

Note: Tomaan is a conceptual design exploration of AI patterns, not a validated machine learning product. It builds on real work I did at Carrene designing offline OTP interfaces for a security-conscious banking user base.