RiskDesk: AI-assisted KYC review designed to clarify fragmented evidence while keeping judgement with the analyst
Role
Product Designer
Scope
Research, Product Strategy, Interaction & UI Design
Domain
TL;DR
RiskDesk Copilot explores how AI can support complex compliance work without replacing human judgement. It helps analysts navigate uncertainty, understand the evidence behind each case, and reach conclusions they can clearly explain and defend.
project overview
RiskDesk Copilot brings customer information, screening results, and supporting documents into a single case view. The concept covers the core review workflow: understanding what triggered an alert, identifying missing or conflicting evidence, and documenting the analyst's decision.
How Might We - Initial framing
How might we help KYC analysts resolve complex onboarding cases faster and with more confidence, while keeping them in control of the final decision?
research Synthesis
An alert does not tell the whole story
A customer may be flagged because their name resembles someone on a risk list, they appear in adverse media, or key company information is missing. None of these signals proves they present a risk.
The analyst must compare sources, assess whether the match is relevant, and reconcile information that may be incomplete, outdated, or contradictory. They must then decide what to do and document the reasoning clearly enough for someone else to review later.
90 to 95%
of screening alerts may be false positives
88%
of firms keep KYC data in two or more systems
What the research revealed
Secondary research showed that KYC review is not a checklist exercise. Analysts interpret incomplete information, investigate ambiguous matches, and document decisions that may be reviewed months later.
High false-positive volumes and fragmented data make this work harder. AI also introduces a new risk: analysts may accept recommendations without enough scrutiny.
The opportunity here was not to automate decisions, but to help analysts gather, understand, and document the evidence behind them.
I thought the problem was speed
My initial assumption was straightforward: analysts had too many alerts, so the product should help them decide faster. Mapping the full workflow revealed something different. Once the context is clear, the decision itself is often quick. Most of the work happens around it: gathering and reconciling evidence beforehand, then documenting the reasoning for review or audit.
This reframed the product. Instead of trying to speed up the decision, RiskDesk would help analysts prepare the case, surface uncertainty, and create a clear record of the outcome. *
01.
Open alert.
02.
Gather context
03.
Interpret
04.
Decide
05.
Document
06.
Close or Escalate
Journey Insight :
The product should reduce the effort of gathering context before the decision and documenting the reasoning afterwards, without making the judgement for the analyst.
Reframed HMW *
How might we reduce the work around complex KYC decisions while keeping judgement and accountability with the analyst?
Who we designed for · KYC operations · Proto-personas
Three roles, one decision chain
Secondary research suggested that KYC review often involves several roles across the decision process.
Primary proto-persona
Maria · First-line analyst
Needs to manage a high volume of cases without missing key signals or writing every rationale from scratch.
Secondary proto-persona
Diego · Senior analyst
Needs escalated cases to show what was reviewed, what remains unresolved, and why.
Secondary proto-persona
Laura · Team lead
Needs consistent, traceable decisions and confidence that AI suggestions were reviewed, not simply accepted.
Why all three roles matter
The core workflow is designed around the first-line analyst, but the decisions they make continue downstream. Escalations need enough context for senior analysts to review, while team leads need decisions to remain consistent and traceable.
Design Implication
Mapping these three roles highlighted a clear tension: Maria needs efficiency, Laura is accountable for quality, and Diego inherits what remains unresolved. RiskDesk had to reduce effort without weakening review.
This shaped a clear product principle: the system gathers, the analyst interprets, and every decision leaves a trace.
COMPETITIVE LANDSCAPE
Existing tools support alert handling. I focused on what happens when the evidence is incomplete or contradictory.
I reviewed six KYC and compliance products across onboarding, case management, screening, fraud, and identity decisioning. I focused on how they present risk, explain recommendations, structure evidence, support escalation, and document decisions. Across the products I looked at, alert handling was well supported, but uncertainty was handled less explicitly.
Established Patterns
Evidence consolidation appeared across all six products reviewed.
WHERE I saw an opportunity
The opportunity I identified was in cases where evidence is incomplete or conflicting: making gaps visible, knowing when AI should withhold a recommendation, and keeping disagreement part of the record.
Visible uncertainty was the least represented pattern in the review.
HOW THE CONCEPT EVOLVED
Evidence before recommendation
The first concept placed a suggested action and a “92% confidence” score at the top of the Copilot. It looked efficient, but risked biasing the analyst toward the recommendation before they reviewed the evidence.
Initial Pattern
Recommendation first
The answer leads. Evidence follows.
A confidence score makes the recommendation feel certain, but does not show what evidence is still missing or conflicting.
Revised Pattern
Evidence first
Signals, sources, and missing information appear before any suggested action.
Additional Safeguard
Visible uncertainty
When the evidence is not enough, the AI can withhold a recommendation instead of forcing an answer.
I replaced the confidence score with three evidence states that describe the quality of the available evidence, not the AI’s certainty. This keeps missing or conflicting evidence visible instead of compressing the case into a single confidence number.
sufficient
missing information
conflicting evidence
When the evidence cannot support a reliable recommendation, the Copilot can return “Unable to assess” instead of forcing an answer.
Validation Focus
This remains a hypothesis. The key test is whether analysts inspect the evidence and challenge the AI even when its suggestion appears convincing.
product walkthrough
From alert to defensible decision
The workflow helps analysts identify what needs attention, investigate the evidence, and document a clear outcome without handing control to AI.
Decision Flow
From evidence to decision
Once the evidence has been reviewed, RiskDesk helps the analyst understand what is complete, missing, or conflicting without turning that state into a decision. The analyst chooses what happens next and, when ready, commits to an outcome and the reasons behind it. Only then does AI step in again, linking those reasons to the evidence already reviewed and helping turn the analyst’s judgement into a clear, editable rationale and traceable record.
01.
Evidence reviewed
02.
ANALYST CHOOSES NEXT STEP
03.
Commit to a decision
FROM DECISION TO RECORD
Consistent records, analyst-owned reasoning
An empty field creates unnecessary manual work. A fully generated explanation creates the opposite risk: approving reasoning the analyst did not fully form or verify themselves.
STRUCTURED REASONS
The analyst selects the outcome and supporting reasons before any rationale is generated.
Editable drafts preserve judgement
RiskDesk links those reasons to the evidence and prepares a draft. The analyst reviews, edits, and owns the final rationale.
The final record keeps the analyst’s decision, supporting evidence, Copilot assessment, rationale, and case history clearly traceable.
conclusion
A hypothesis, not a result
RiskDesk has not yet been tested with KYC analysts. The next step is to test whether showing evidence before an AI suggestion helps analysts review cases more critically without adding unnecessary friction. These are the first things I would test.
01.
02.
03.
04.
05.
Key Question
If analysts still accept the AI suggestion without reviewing the evidence, is changing the order enough, or should they form an initial judgement first?