Back to Selected Work

CASE STUDY · MORGAN STANLEY

Intelligent Document Processing

AI-Assisted Document Validation Workflow

Embedding intelligent document validation into existing financial workflows while preserving control, trust, and operational continuity.

AI/MLWorkflow IntegrationEnterpriseHuman-in-the-Loop

ROLE

Principal UX Designer

USERS

Operations analysts, compliance officers, document review teams

PLATFORM

Enterprise web platform

FOCUS

AI trust, human-in-the-loop verification, confidence scoring, exception handling

TOOLS

Figma, prototyping, user research, workflow mapping

CONFIDENTIALITY NOTE

Some details, screens, and workflows are recreated or generalized to protect confidential information while preserving the design challenge, process, and decision-making approach.

01
SITUATION

Situation

Business context

Manual document processing was a major bottleneck across compliance workflows, prone to human error and carrying high operational costs.

User context

Operations analysts needed to review large volumes of documents daily — verifying data extraction accuracy while meeting strict processing SLAs.

System context

AI extraction models were available but untrusted. Users had no visibility into confidence levels or the reasoning behind automated decisions, so they defaulted to full manual review.

02
DESIGN CHALLENGE

Design Challenge

Design an interface where humans can verify AI-extracted data with high confidence and minimal friction — building trust in automation while preserving the control and accuracy required in financial compliance.

Low trust in automation

Users defaulted to full manual review because the AI provided no evidence for its conclusions — the system felt like a black box.

High error cost

Incorrect data extraction in financial documents carries regulatory and financial risk. Users could not afford to blindly accept automated results.

Processing bottleneck

Manual review speeds could not keep pace with increasing document volume and tightening compliance requirements.

AI confidence transparencyRegulatory complianceHigh-volume processingZero-error toleranceAudit trail requirements
03
MY ROLE

My Role

What I led

The human-in-the-loop verification interface design and the AI feedback mechanism that allows analyst corrections to improve the model over time.

What I designed

Side-by-side document comparison views, field-level confidence scoring UI, inline editing patterns, exception handling flows, and the audit trail interface.

Who I partnered with

ML engineers, compliance officers, operations managers, front-line document analysts, and the engineering team building the extraction pipeline.

04
MAKING THE WORKFLOW VISIBLE

Making the Workflow Visible

Mapped the transition points between automated extraction and manual verification to identify the high-risk handoffs where human judgment is most critical — and where trust breaks down.

DOCUMENT PROCESSING WORKFLOW · AI EXTRACTION → CONFIDENCE SCORING → HUMAN VERIFICATION → FEEDBACK LOOP
05
EXPLORING THE EXPERIENCE

Exploring the Experience

Confidence visualization

Explored visual treatments for AI confidence scores at both the field level and the full document level.

Side-by-side comparison

Tested layouts for comparing the source document directly against extracted data fields.

Exception handling flows

Designed the flows for flagged fields, escalation paths, manual override, and supervisor review.

Batch processing patterns

Explored patterns for high-volume document review across priority-sorted queues.

Feedback loop mechanism

Prototyped the mechanism for analyst corrections to feed back into the AI model as training data.

Audit trail design

Designed the verification history view for compliance documentation and regulatory reporting.

06
THE SOLUTION

The Solution

A unified workspace where AI suggestions are integrated directly into the existing document review flow — with transparent confidence scoring, inline correction capabilities, and a feedback loop that continuously improves the model.

PRIMARY SOLUTION · DASHBOARD VIEW

Document queue

Priority-sorted queue with AI confidence indicators, SLA tracking, and batch processing controls.

Extraction view

Side-by-side comparison of the source document and extracted fields with field-level confidence scores.

Confidence scoring

Visual indicators showing AI certainty at both field and document level, guiding reviewer attention to uncertain areas.

Inline correction

Direct editing of extracted values with automatic AI model feedback — every correction improves future accuracy.

Exception dashboard

Centralized view of flagged documents requiring manual review, supervisor escalation, or compliance attention.

Audit history

Complete trail of all verifications, corrections, and approvals for regulatory compliance and reporting.

07
KEY DESIGN DECISIONS

Key Design Decisions

Highlight-and-verify model

WHYFull re-entry wastes time. Highlighting only uncertain fields focuses reviewer attention where it matters most.
TRADEOFFUsers must trust the AI's pre-fill for non-highlighted fields — but confidence scores make this trust informed.
RESULTFaster processing without sacrificing accuracy on the fields that carry the most risk.

Transparent confidence scores

WHYTrust requires evidence. Showing exactly why the AI is uncertain builds informed reviewer confidence.
TRADEOFFMore visual complexity in the interface vs. better-informed decision-making by analysts.
RESULTAnalysts reported higher trust in the system and made faster, more confident review decisions.

Integrated feedback loop

WHYEvery analyst correction represents training data. The UI should be a surface for continuous model improvement.
TRADEOFFAdditional complexity for analysts vs. long-term gains in AI accuracy and reduced review burden.
RESULTContinuous model improvement driven by real analyst interactions, reducing the need for manual review over time.

Progressive automation tiers

WHYDifferent document types carry different risk levels — the level of human oversight should match the risk.
TRADEOFFMore configuration and automation logic vs. appropriate, risk-matched human oversight.
RESULTLow-risk documents processed with minimal friction while high-risk documents received full human review.
08
IMPACT

Impact

USER IMPACT

Analysts shifted from full manual review to targeted, confidence-guided verification — focusing attention where it matters most and reducing cognitive burden.

BUSINESS IMPACT

Significantly increased document processing capacity while maintaining the compliance standards required in financial services.

OPERATIONAL IMPACT

Reduced the manual processing bottleneck, improved SLA adherence, and created a path for increasing automation over time.

AI TRUST IMPACT

Established a pattern for transparent, human-centered AI integration that influenced the approach for other enterprise AI initiatives.

09
REFLECTION

Reflection

What I learned

Designing for trust in AI is less about perfect automation and more about providing clear, transparent evidence for the AI's conclusions — users trust what they can understand.

What I would improve

I would explore more granular control over automation levels, allowing experienced analysts to customize their own review thresholds based on their confidence and expertise.

How this shaped my approach

This project established my framework for human-AI collaboration: transparency first, automation second, user control always. The human must always be able to understand why the AI made its decision.