RiskOS

The AI prepares, the analyst decides

RiskOS is a working fraud-detection prototype. It tests whether an AI can prepare alerts while the analyst keeps the decision, with reasoning written out in full and a visible trace for every action.

Role: Product design
Scope: Agentic UX experiment
Format: Working prototype

View prototype

Summary

Context: When an AI takes part in a decision, how do you keep the person deciding actively engaged? I chose fraud detection in European banking to explore the question: analysts work under time pressure, 80% of their alerts turn out to be false alarms, and the final decision is always theirs.
My role: A personal exploration. I designed the interaction model and built the working prototype: React 18, Vite and Tailwind CSS, deployed on Vercel.
Deliverables: Triage view
Five cases sorted by risk level, with priority filters and a live counter.
Streaming AI analysis
The AI writes its reasoning word by word, with highlighted details, sources and a confidence score.
Decision flow
Block the card, escalate or monitor, with the Slack message and the customer SMS made visible.
False-alarm resolution
A medium-risk alert reviewed by the AI and confirmed with one click.
Queue processing
Five cases resolved in sequence, with a session recap.

An AI that helps without disengaging the analyst

The challenge

If the AI does too much, the analyst disengages. If it does not explain itself, trust erodes. In many teams the daily tool is still a spreadsheet and fixed rules, with no context, no sorting by relevance and no memory of earlier cases. An analyst at a European neobank handles 80 to 150 alerts per shift, and every minute spent on a false alarm is a minute not available for a real case.

Same data, two readings.

The objectives

The prototype tests whether the AI can handle the preparation work while the analyst keeps full ownership of the decision. Three principles guided the design. The AI sets up the context and the analyst decides. The reasoning is written out in full, so the analyst can read it. When the analyst acts, the effect is visible outside the tool as well.

In the process, a suspicious transaction first meets the bank’s automated rules. If it is flagged, the alert enters a queue, the AI analyzes it, and the analyst reviews and decides.

Where RiskOS sits in the process.

Three design decisions

Triage under time pressure

The analyst opens a session and finds five cases sorted by risk level, with color-coded priorities, a filter and a live counter. Without sorting, the analyst scrolls the whole list to find the urgent cases. Here, sorting takes a few seconds.

Triage view

Writing the AI reasoning in real time

The AI writes its analysis word by word. Relevant details are highlighted as they appear, the data sources light up progressively and a confidence score shows the level of certainty. The action buttons stay hidden until the analysis is complete, so the analyst reads the full reasoning before being able to decide.

AI analysis, streaming

Making the effect of each action visible

The analyst picks an action: block the card, pass the case to a senior or keep it under watch. A confirmation screen recaps what happened. Then two things appear that usually stay invisible: the Slack message to the fraud team and the SMS to the customer. Usually the analyst acts and never sees the result.

Confirmation and what happened next

The prototype in use

Resolving a false alarm in eight seconds

A medium-risk alert arrives: score 45, a 450 euro payment. The AI reviews the transaction history and finds nothing unusual. The analyst confirms with one click.

False alarm, resolved

Processing a full queue

The analyst works through five cases in sequence. After each resolution, the next case loads automatically. At the end: five cases resolved, 92 seconds total.

Case flow and session recap

What the project shows

Streaming the reasoning builds trust in a way scores cannot.

When the AI writes its analysis word by word, the analyst reads along and forms their own view at the same time. A confidence score after the fact just says “trust me” without showing the work.

Hiding the buttons until the analysis is done changes how people read.

Under pressure, people click the first thing available. Holding the decision buttons back until the AI finishes writing gives the reasoning a chance to land.

What remains to be checked

The prototype runs on five fictional cases and produces no measurement: its effect on the decisions of real analysts remains to be verified. The same dynamics exist in compliance, medical triage, content moderation and incident response.

GitHub

Victor Soussan

Lead Product Designer

Interested in working on a similar project? Let's talk.

Contact LinkedIn

The AI prepares, the analyst decides

Summary

An AI that helps without disengaging the analyst

The challenge

The objectives

Three design decisions

Triage under time pressure

Writing the AI reasoning in real time

Making the effect of each action visible

The prototype in use

Resolving a false alarm in eight seconds

Processing a full queue

What the project shows

Streaming the reasoning builds trust in a way scores cannot.

Hiding the buttons until the analysis is done changes how people read.

What remains to be checked

Toolkit

Tessan

Dailymotion Partner

The AI prepares, the analyst decides

Summary

An AI that helps without disengaging the analyst

The challenge

The objectives

Three design decisions

Triage under time pressure

Writing the AI reasoning in real time

Making the effect of each action visible

The prototype in use

Resolving a false alarm in eight seconds

Processing a full queue

What the project shows

Streaming the reasoning builds trust in a way scores cannot.

Hiding the buttons until the analysis is done changes how people read.

What remains to be checked

Summary

An AI that helps without disengaging the analyst

The challenge

The objectives

Three design decisions

Triage under time pressure

Writing the AI reasoning in real time

Making the effect of each action visible

The prototype in use

Resolving a false alarm in eight seconds

Processing a full queue

What the project shows

Streaming the reasoning builds trust in a way scores cannot.

Hiding the buttons until the analysis is done changes how people read.

What remains to be checked

Keep reading

Toolkit

Tessan

Dailymotion Partner

Summary

An AI that helps without disengaging the analyst

The challenge

The objectives

Three design decisions

Triage under time pressure

Writing the AI reasoning in real time

Making the effect of each action visible

The prototype in use

Resolving a false alarm in eight seconds

Processing a full queue

What the project shows

Streaming the reasoning builds trust in a way scores cannot.

Hiding the buttons until the analysis is done changes how people read.

What remains to be checked