Document Intelligence
Designing clarity and trust across complex financial documents
ROLE
UX Designer III
TIMELINE
2023 - Present
PLATFORM
Enterprise web application
PRIMARY USERS
Investment professionals
FOCUS AREAS
Cross-Document Intelligence • Analytical Workflows • Verifiable AI Systems
Helping investment professionals quickly find, interpret, and compare insights across large volumes of high-stakes financial documents.
CONTEXT
WHAT IS DOCUMENT INTELLIGENCE?
Document Intelligence is the primary workspace investment professionals use to search, read, analyze, and extract insights from filings, transcripts, research, and news. As AI capabilities expanded, they became increasingly integrated throughout that workflow.
WHAT IS THE USER’S REALITY?
Users make time-sensitive, high-stakes decisions under constant pressure, working through long, dense documents that must be scanned, compared, and interpreted across companies, time periods, and sources.
WHY IS THIS PROJECT IMPORTANT?
As AI features became more embedded into the product, the bar moved. Speed was essential, but what really mattered was whether users could trust what they were seeing.
AI summaries, cross-document synthesis, sentiment signals had to feel grounded. If insights weren’t clearly tied back to the source, even small ambiguities could chip away at confidence.
The evolution wasn’t about speed. It was about designing confidence into insight.
WHAT WILL THIS CASE STUDY COVER?
This case study focuses on key moments where design decisions helped users move faster without sacrificing confidence or accuracy. It shows how trust was intentionally designed into AI-assisted workflows instead of being treated as an afterthought.
KEY CONSTRAINTS
Content
• Large, text-heavy documents
• Complex financial language
• Multiple document types
Workflow
• Time-sensitive decisions
• Cross-document comparison
• Non-linear analysis patterns
Trust
• High expectations for traceability
• AI outputs must withstand scrutiny
• Confidence degrades without grounding
Primary Users
Investment Management
Junior Analyst
Earnings Calls
Filings
Portfolio Tracking
Investment Banking
Junior Analyst
Deal Research
Comparables
Due Diligence
Private Equity
Junior Associate
Target Screening
Sector Analysis
Investment Committee Memos
THEIR REALITY
Junior analysts/associates are responsible for doing the bulk of document-heavy analysis that informs high-stakes investment decisions. To support portfolio managers and investment teams, analysts must extract key signals, validate assumptions, and compare insights across companies and time periods often from long, dense documents.
VOLUME
Working across multiple long-form documents simultaneously; transcripts, filings, research, news
PRESSURE
Synthesizing information quickly under scrutiny from senior stakeholders with little margin for error
ACCOUNTABILITY
Moving fast while maintaining defensible, source-backed accuracy. The outputs must withstand challenge
While senior investors consume the outputs, junior analysts do the heavy document work.
Problem
Investment professionals could access vast amounts of information, but struggled to identify what mattered, synthesize insights across sources, and validate AI-generated outputs against the underlying evidence.
As Document Intelligence evolved beyond search into AI-assisted analysis, a new challenge emerged. Users needed to trust AI-generated insights, not just consume them.
That challenge surfaced in three recurring design tensions:
INFORMATION DISCOVERY
• Insights buried in documents
• Linear navigation slowed scanning
• Difficulty jumping to key moments
COMPARISON AND SYNTHESIS
• One-document-at-a-time workflows
• Manual mental stitching across sources
• No persistent analytical context
TRUST IN AI INSIGHTS
• Summaries lacked context
• Needed explicit sourcing and traceability
• Low transparency reduced confidence
DESIGN QUESTION
How might we design workflows that accelerate insight without compromising traceability, analytical rigor, or user trust?
Approach
Designing for Speed Without Compromising Trust
To help investment professionals move faster without sacrificing transparency or trust, I helped shape the Document Intelligence experience around three strategic design principles that guided our product decisions and future AI workflows:
CLARITY THROUGH LAYERED INSIGHT
• Structure complexity into progressive layers of insight
• Surface the most relevant signal first
• Let users drill down on their own terms
CROSS-DOCUMENT INTELLIGENCE AT SCALE
• Enable thesis-driven research across documents
• Surface patterns across sources while preserving traceability
• Anchor analysis to a bounded, intentional scope
REUSABLE EVIDENCE NAVIGATION
One interaction model reused across AI experiences
Every insight connected to supporting evidence
Trust built through consistent interaction patterns
This wasn't a collection of features. It was a design approach that balanced speed, trust, and scalability across every AI experience.
CHAPTER ONE
Clarity Through Layered Insight
Reduced Scan Time While Preserving Interpretability
Investment analysts don’t read documents the way they were written. They read them to extract what matters under time pressure. Earnings calls, filings, and research reports are long, dense, and structured chronologically. But analyst questions are not chronological. They are thesis-driven:
"Did guidance change?"
"Are there new risk signals?"
"Is tone shifting?"
"What actually matters here?"
The platform gave analysts access to information, but not a way to quickly orient themselves around those questions. Search reflected the document's chronology rather than the analyst's intent. Sentiment surfaced high-level signals without revealing the evidence behind them. AI summaries accelerated reading, but didn't make their conclusions traceable.
Insight existed but it remained fragmented, forcing analysts to piece together the story themselves.
Section 1 of 3
Finding Mentions, Not Meaning
THE CHALLENGE
Search Results Were Organized by Chronology, Not Relevance
Search results followed the order of the document, even though analysts approached research with specific questions in mind. Users could find keyword matches, but still had to manually interpret which mentions were material and which were noise.
DESIGN RESPONSE
From Chronology to Relevance
Rather than asking analysts to read every match equally, we redesigned the experience to progressively surface the signals most likely to answer their question. Users could still switch to chronological view, preserving control, but the default experience aligned with the user's intent. This reduced the time spent scanning low-value mentions and helped users quickly assess whether a document warranted deeper review.
Proto: Relevance-First Exploration
Results were ranked by analytical importance rather than chronology, but sorting by chronological order was only one click away
Section 2 of 3
Sentiment Without Context
THE PROBLEM
Sentiment as a Detached Metric
Sentiment was presented as historical trend graphs in a side panel that included net positivity, numerical transparency, and language complexity. While useful for benchmarking against industry averages, the experience lacked context. Users could see sentiment trends over time, but they couldn't inspect the passages that explained them.
DESIGN RESPONSE
Contextual Sentiment Within the Document
We transformed sentiment from detached metrics into a contextual reading layer. Sentiment signals were grouped by theme and organized by importance (high, medium, low). These varying importance levels allowed analysts to see tonal shifts tied to the exact language driving them. Sentiment became a navigable signal map across the document.
To keep that map usable, we highlighted only the highest-impact positive and negative signals by default. This directed attention to the moments most likely to influence an analyst's thesis, while reducing visual noise. When deeper investigation was needed, analysts could progressively expand into medium- and low-importance excerpts without losing context or control. Sentiment became a tool for investigation rather than measurement.
We kept the net positivity trend chart, but repositioned it as supporting context, while contextual sentiment became the primary way analysts explored the document.
Proto: Progressive Sentiment Exploration
Analysts began with the strongest signals and progressively expanded into supporting excerpts, balancing clarity with depth.
Try the proto
TENSION
Should Sentiment Live on the Right or Left Panel?
Sentiment as a Metric vs. Sentiment as Navigation
In early versions of Document Intelligence, sentiment lived in the right-hand panel alongside other analytical outputs. It was presented as something analysts evaluated, not something they navigated with.
As sentiment became embedded directly within the transcript, analysts began using it differently. It was no longer something to evaluate, it became a way to navigate directly to the passages that mattered.
That shift led us to relocate sentiment controls to the left-hand panel alongside search and filtering. Because it was an established feature, stakeholders initially questioned the move. Once we demonstrated that sentiment now supported navigation instead of retrospective analysis, the rationale became clear. The new placement better matched how analysts explored documents.
By repositioning sentiment as navigation, we aligned the interface with how analysts actually explored documents, not just how they evaluated them.
Section 3 of 3
AI Summaries Without Traceability
THE PROBLEM
Summaries as Detached Outputs
AI summaries were positioned at the top of the document and generated on demand. They provided a quick overview of key themes, but operated independently of the transcript itself. Users could read the conclusions, but could not trace them back to supporting excerpts. Validation required manual scanning.
DESIGN RESPONSE
Make Every Insight Inspectable
Instead of treating AI summaries as static outputs, we turned them into entry points into the underlying evidence.
Each summary point became directly connected to the transcript. Clicking an insight immediately navigated analysts to the supporting passage, allowing them to verify claims without manually searching through the document.
Because most insights were supported by a single excerpt, we avoided unnecessary visual clutter. Only insights with multiple supporting passages displayed a navigation indicator, allowing analysts to step through each supporting excerpt one at a time. The result was traceability that stayed lightweight—present when needed, invisible when it wasn't.
Proto: Traceable Summaries
Single-source insights navigated directly to evidence. Only insights supported by multiple excerpts displayed lightweight navigation
Try the proto
TENSION
Should AI Summaries Be the Default Entry Point?
Default Summaries vs. Conversational Exploration
We debated which experience should anchor the document workflow: pre-generated summaries or query-based exploration through ChatIQ. Making summaries the default would have positioned AI as the starting point for analysis, encouraging passive consumption instead of analyst-led investigation.
We chose the opposite approach. ChatIQ became the default entry point, encouraging users to begin with questions rather than answers. AI summaries remained one click away as a secondary tab, helping users quickly orient themselves without replacing the analyst's own line of inquiry.
TAB ORDER
Chat remained the default entry point. Summary stayed one click away—available for orientation without replacing analyst-led inquiry.
CHAPTER TWO
Enabling Cross-Document Intelligence at Scale
Shifted From Document Reading to Thesis-Driven Synthesis
Clarity within a single document was only the first step. Investment analysts rarely evaluate earnings calls in isolation. They compare companies, track themes across quarters, and connect evidence across an entire body of research.
As AI capabilities expanded, the challenge shifted from understanding one document to helping analysts answer questions across many. The challenge became: How do we connect insights across documents without losing context?
Analysts came to the platform with questions like:
What changed this quarter?
Which companies are discussing AI infrastructure?
How is sentiment shifting across competitors?
What themes are emerging across sectors?
None of these questions could be answered within a single document. Rather than treating documents as isolated containers, we began designing for continuity across research workflows, enabling analysts to move fluidly from discovery, to comparison, to synthesis without losing their place.
The goal wasn't just speed. It was continuity.
SYSTEM BOUNDARY
Designing Within Constraints
Early versions of the system were limited to analyzing up to 20 documents at a time. Rather than treating this technical constraint as a temporary limitation, we asked how it could improve the analytical workflow.
Instead of encouraging users to analyze everything at once, we used the limit to reinforce intentionality. The cap encouraged analysts to define a working set: a focused collection of documents assembled around a specific analytical question.
This reinforced one of our core principles: AI should help analysts investigate what matters, not attempt to process everything. By encouraging users to define a deliberate working set, the experience stayed focused, explainable, and grounded in a clear scope.
Designing for Investigation
Analysts do not work linearly through documents. They move between discovery, selection, analysis, and validation—continuously refining their understanding as new evidence emerges.
Rather than forcing a linear workflow, the experience supports four essential aspects of investigation:
01 • Thesis-driven workflows
Search
Start with a question
Analysts investigate questions across collections of documents rather than reading each one in isolation. Multi-document chat was designed to support that workflow.
The result is a workflow that stays centered on the analytical question instead of any single document.
02 • Working Set
Select
Define the working set
We introduced the concept of a working set: a curated group of documents tied to a specific analytical question.
Analysts define scope upfront and carry that context into a persistent Multi-doc chat session, ensuring every response remains grounded in a known body of evidence.
03 • Analytical continuity
Analyze
Preserve continuity across surfaces
Analysts rarely work in a straight line. They constantly move between searching, reading, comparing, and validating information. We preserved continuity across these transitions so analysis could continue without losing momentum.
04 • Grounded intelligence
Validate
Keep every insight traceable
Every AI response remained connected to its supporting evidence through direct links back to the source material.
AI wasn't treated as a source of truth. Every response remained connected to the evidence behind it, allowing analysts to inspect, validate, and challenge each conclusion.
PRESERVING ANALYTICAL CONTINUITY
Supporting Continuous Investigation
Defining a working set was only the beginning. Analysts rarely stop exploring once they begin investigating. New findings often send them back to refine filters, inspect additional documents, or assemble another collection of evidence.
The challenge was deciding how Multi-document Chat should fit into that workflow. Should investigation happen alongside discovery, or become its own dedicated workspace? Early concepts kept both activities together, but combining discovery and synthesis created a crowded interface where exploration and analysis competed for attention.
Instead, we treated Multi-document Chat as a dedicated investigation workspace. Analysts could continue discovering documents while preserving the context of the analysis already in progress. Each conversation maintained its own working set, making it easy to compare evidence, return to discovery, or begin a new investigation without losing analytical continuity.
Proto: Opening Multi-Document Chat Within the Same Workspace
An early concept, later replaced by a dedicated workspace
TENSION
Why not let AI automatically retrieve the most relevant documents?
AI-assisted retrieval vs analyst-defined scope
Modern AI systems can retrieve relevant documents and generate answers in seconds. But for investment research, speed wasn't the hard problem. The real question was: Who defines the evidence behind the answer?
If AI automatically assembled the document set, analysts would lose visibility into what was included, what was left out, and why. Without a clearly defined evidence set, answers become harder to validate, compare, and defend.
We chose a different approach. Analysts intentionally defined the working set, and AI reasoned within those boundaries. Every response remained grounded in a transparent, reproducible set of evidence.
AI-DRIVEN RETRIEVAL
• Faster to start
• AI defines the evidence set
• Selection criteria less visible
• Results may vary as retrieval changes
ANALYST-DEFINED SCOPE
• Analyst controls the evidence set
• Stable, reproducible analysis
• Every insight tied to known inputs
• Easier to validate and defend
Rather than asking AI to decide what mattered, we designed it to help analysts make sense of the evidence they intentionally selected.
Putting It Together
An End-to-End Analyst Workflow
Define scope
Select a working set of documents (max 20)
Launch ChatIQ
Initiate a session grounded in these documents.
Ask questions
Ask questions that surface patterns.
Review
Responses surface patterns that can be traced back to source.
Validate
Jump directly into supporting excerpts to inspect evidence.
CHAPTER THREE
Designing a Reusable Interaction Model
One Transparency System Across Summaries, Chat, Sentiment, and Multi-Document Analysis
Connecting AI insights to evidence established trust, but it also exposed a new design problem. As AI responses became more sophisticated, a single insight was rarely supported by a single passage. Analysts needed a way to inspect increasingly complex evidence without disrupting the flow of reading.
This chapter explores how a transparency pattern first introduced in AI Summaries became a reusable system across conversational AI. Rather than inventing a new citation model for each feature, we refined and reused one interaction model that scaled from single excerpts to multiple passages and, eventually, multiple documents.
THE PROBLEM
When One Insight Came From Many Sources
Our first implementation linked every summary insight to a single supporting excerpt. It established traceability, but it assumed each insight could be explained by one passage. As AI responses became more sophisticated, that assumption quickly broke down. Many insights synthesized information from multiple excerpts throughout a transcript, making one-to-one citations increasingly insufficient. The challenge wasn't simply showing evidence, it was helping analysts understand how multiple pieces of evidence supported a single conclusion.
DIRECT EVIDENCE NAVIGATION
Clicking a summary insight took analysts directly to the supporting passage in the document.
THE DESIGN EVOLUTION
Moving Beyond One-to-One Citations
As ChatIQ evolved, sourcing became significantly more complex. Unlike AI summaries, conversational responses could reference multiple excerpts within a document, requiring a more flexible approach to evidence navigation.
We explored several ways to connect AI insights to their supporting evidence, including text highlighting, inline links, pagination controls, and embedded navigation. Each approach solved part of the problem, but none clearly communicated the relationship between an insight and its supporting evidence. They introduced visual noise, interrupted reading flow, or required users to learn new interaction patterns.
Design Explorations That Didn't Scale
We tested multiple interaction patterns before arriving at the final solution.
Too Many Controls
Combining navigation links, icons, and underlining created unnecessary visual complexity. Users had to interpret the interface before they could interpret the evidence.
Looked Like Text Selection
The navigation already brought users to the correct passage. The additional highlighting looked like text selection and introduced visual noise without adding meaningful context.
Unclear Relationship to Evidence
Icon-only links were ambiguous. Users couldn't tell whether they would open a link or navigate to supporting evidence. The approach also didn't scale, since insights could be supported by up to ten excerpts, resulting in a row of icons that quickly became visually cluttered.
Reusing Existing Source Navigation
We reused the down-arrow pattern from other parts of the product to indicate supporting evidence, while separate arrows let users navigate between multiple excerpts. It didn't work because the two arrows made it feel busier and less intuitive than intended.
The breakthrough came when we stopped treating citations as references and started treating them as lightweight navigation. Rather than introducing a new interaction for ChatIQ, we returned to the evidence navigation pattern that had already proven successful in AI Summaries.
We had initially ruled it out because conversational responses contained citations much more frequently than summary bullets, raising concerns about layout stability and visual clutter. After exploring multiple alternatives and working through the technical constraints, we realized the original interaction model was still the clearest solution. Instead of inventing something new, we preserved one consistent mental model across both experiences.
Proto: Adaptive Evidence Navigation
Single-source insights navigated directly to evidence, while multi-source insights introduced lightweight controls for stepping through supporting excerpts. This final pattern adapts to the complexity of the insight and stays out of the way until needed
OUTCOME
One Transparency System Across Every AI Feature
What started as an evidence navigation pattern in AI Summaries ultimately became a reusable transparency system across our AI experiences. Rather than creating a different sourcing model for each feature, we preserved one interaction pattern that users could recognize wherever AI generated insights appeared.
Instead of designing citations three different times, we designed one transparency system that every AI feature could inherit. Whether analysts were reading AI summaries, conversational responses, sentiment insights, or multiple documents, supporting evidence behaved the same way. The mental model remained consistent even as the complexity of the underlying data increased.
The result was a more scalable interaction system that reduced implementation complexity, created a familiar experience across AI features, and made generated insights easier to understand, validate, and confidently share.
Multi-Document Chat Sourcing
The navigation model expanded from one transcript to many. Analysts can step through supporting passages across multiple documents using the same iteration
Sentiment Sourcing
The same pattern connects sentiment insights directly to the language behind them, making sentiment transparent and inspectable
Instead of designing citations three different times, we designed one transparency system that every AI feature could inherit.
Outcomes
The Document Intelligence redesign didn't just improve how analysts review documents, it became an adopted research destination and foundation for AI-driven workflows at Capital IQ Pro.
1
From an Underused Viewer to a Core Research Destination
The redesigned Document Intelligence experience transformed an underused document viewer into a heavily adopted research surface. The new viewer reached ~20K weekly unique users during peak weeks. Documents also grew to roughly 4K weekly users from top navigation after becoming a primary destination.
2
AI Became Part of the Workflow
Between November and February, 12.6K users engaged with ChatIQ. Most began with suggested questions, creating an approachable entry point into conversational research while still allowing analysts to move into their own queries as their investigation deepened.
3
Customer Response Validated the Direction
Feedback following launch was overwhelmingly positive, with customers specifically recognizing the value of combining information retrieval with AI-generated insight. Continued enhancement requests—including multi-document ChatIQ—also helped shape the next generation of the experience.
“What you are doing is saving time by gathering data AND providing insights.” – Analyst
Reflections
What I Learned
This project reinforced how much impact can come from rethinking the foundation of an experience rather than simply adding features to it. What started as an underused Document Viewer became a place where analysts could search, explore, and use AI to understand information more efficiently. I learned to think beyond individual features and design patterns that could grow with the product—while keeping clarity, trust, and the source document at the center.
Impact as a Designer
My role grew beyond designing the initial experience. I helped establish the UX foundation for Document Intelligence and continued shaping how new capabilities fit into the broader workflow. That meant translating evolving product requirements, customer feedback, usage data, and technical constraints into experiences that felt connected rather than like a collection of features. As the product expanded, I became increasingly focused on the system as a whole: how analysts move from discovering information to understanding it, validating it, and deciding where to go next.
What’s Next
Document Intelligence has continued to evolve well beyond the initial launch. More recent work has explored Trending Topics, richer search visualizations, and workflows for uploading and analyzing proprietary documents, along with continued improvements across search, discovery, and AI-assisted research. Each new capability builds on the same foundation: helping analysts surface meaningful information faster while keeping them connected to the documents and evidence behind it.