Back to Work
Case Study 01

Turning AI Into a Co-pilot Auditors Trust

GenIAus — EY's GenAI auditing assistant, rebuilt from an ignored feature into a co-pilot auditors actually rely on.

ClientErnst & Young
RoleUX Designer · Lead
TeamPM · PO · 10 Developers
GenIAus — Virtual Internal Auditor home screen

GenIAus was built by engineers. It showed.

The Problem

A GenAI audit tool EY had already shipped — that auditors avoided because it gave them no hierarchy, no onboarding, no way to learn it.

My Approach

Reframed it from a UI cleanup into a systems-level redesign — 8 decisions, shipped incrementally with a 10-developer team.

GenIAus was live, but nobody could figure out how to use it.

  • No hierarchy — a global feature buried in a local sidebar
  • No onboarding, no prompts — no visibility into what the AI could do
  • Trial and error — in a regulated, high-stakes audit workflow
  • Adoption was stalling — EY had already invested, audit teams weren't using it

A rescue operation, not a greenfield project.

Eight states from the engineer-built beta, annotated during contextual analysis — every Nielsen heuristic violated, each violation compounding the next. Step through each frame to see the failure mode.

Every constraint stacked against a clean rebuild — this had to be a rescue, not a redo.

Systems thinking, not screen design.

I re-framed GenIAus from a collection of AI features into a risk-aware human–AI decision system — built on reusable interaction patterns, contextual AI behaviors, and embedded transparency.

This wasn't a UI cleanup — the problems were structural, so the fix had to be too.

01 Layer 01 End-to-end workflow redesign 02 Layer 02 Context-aware AI interactions 03 Layer 03 Human–AI collaboration model 04 Layer 04 Scalable IA & design patterns 05 Layer 05 Ethics & accessibility as system rules Each layer reinforces every other
Fig 09The systems thinking framework I used to approach GenIAus — five interconnected layers that ensured every design decision reinforced the whole, not just individual screens.

In regulated domains, users won't adopt tools they can't explain to a regulator. Every design decision had to answer the question: "Could an auditor justify this to their oversight board?"

Input New workflows, prioritized features, & design system. Translated through systems- level redesign into five measurable outcome pillars. Outcomes · 5 pillars 01 User-friendly interface Approachable for non-technical users; built around how audit and risk teams actually work. 02 Time-saving efficiency Smart Quickstart and prompt scaffolding cut analysis time from hours to minutes. 03 Highest standard of data security & privacy Internal-only deployment; persona-scoped access; transparent data lineage. 04 Contextual understanding Persona, prompt, and document grounding so outputs reflect real audit context. 05 AI-powered document analysis Inline summaries, evidence extraction, and cross-document synthesis at scale.
Fig 10Transforming GenIAus for Success — mapping new workflows, prioritized features, and the design system to five outcome pillars: user-friendly interface, time-saving efficiency, data security compliance, contextual understanding, and AI-powered document analysis.

Five days from analysis to clickable prototype.

A structured 5-day design sprint moved the work from analysis to validated structure — divergent before convergent, lo-fi before hi-fi, structure before surface. The carousel below cycles through the sprint's working materials; it advances on its own, but pause to study a frame any time.

5 days, from teardown to a clickable, testable prototype.

Day 1–2
Understand & Define

Contextual analysis + stakeholder interviews defined the core problem.

Day 3
Ideate

Flow variations on Miro, testing structure vs. flexibility.

Day 4
Decide

Evaluated on time-to-value, scalability, buildability.

Day 5
Prototype

Lo-fi to validate structure, then hi-fi in Figma.

"How might we redesign GenIAus so that its AI capabilities are intuitively understood and seamlessly integrated into the auditor's existing workflow?"

Four insights that changed the design.

After the sprint, I led usability testing with audit professionals — the actual end users — using a moderated think-aloud protocol focused on Document Intelligence and Prompt Scaffolding. Insights were synthesized through thematic analysis, empathy mapping, and affinity mapping.

Users couldn't act on what they couldn't see or understand.

Labeling mattered more than expected

"Find in Files" didn't say what it did. Renaming to "Document Intelligence" fixed comprehension immediately.

Users needed to see capabilities first

The empty chat state was a dead end. Structured entry points gave users a mental model before they typed a prompt.

File selection needed progressive disclosure

80+ files in a flat list overwhelmed users. An audit-name dropdown first, then a filtered list, cut decision fatigue.

Trust required transparency

The AI's limitations had to be stated upfront — not legal cover, but what made people willing to engage.

Eight features. Eight specific problems.

Each redesigned feature was a deliberate response to a specific insight — the problem, the design decision, and the reasoning behind it.

8 problems, 8 targeted fixes — the problem, the decision, and why it works.

01
Smart Quickstart — getting users productive without confusion.
AccessibilityReduce cognitive load

The Problem

  • Empty chat, no guidance
  • Low-literacy users didn't know where to start
  • High-literacy users didn't know what this AI could do

The Decision

  • Landing state that doubles as onboarding + navigation
  • 4 capability entry points, each with icon + label
  • Ask me anything, Document Intelligence, Q&A in Issue Tracker, Q&A in Analytics
  • Transparency disclaimer sets AI-limitation expectations below

Why It Works

  • Answers "what is this, what can it do, what do I do first"
  • No documentation or tutorial required
Smart Quickstart — landing state with four capability entry points
Fig 17Smart Quickstart — the redesigned landing state. Four clear capability entry points replace the blank chat, and the disclaimer below sets expectations about AI limitations upfront.
02
Prompt Scaffolding — making AI interactions purposeful.
AccessibilityReduce reliance on prior knowledge

The Problem

  • Users didn't know how to prompt the AI effectively
  • Enterprise users aren't ChatGPT power users — needed guidance

The Decision

  • Persistent contextual prompt suggestions at the bottom of chat
  • Tailored to active feature mode, updates with context
  • Tap a suggestion to start, then refine

Why It Works

  • Turns a blank text field into a guided interaction
  • Barrier drops from "compose from scratch" to "select and customize"
Prompt scaffolding — contextual suggestion chips beneath the input
Fig 18Prompt scaffolding in action. Contextual suggestion chips beneath the input field reduce the blank-page problem and teach users the system's language through use.
03
Conversation History — continuity and confidence.

The Problem

  • No memory — every session started fresh
  • Multi-week engagements lost context between sessions
  • Auditors forced to repeat queries

The Decision

  • Collapsible sidebar, last 10 chats with preview text
  • Resume any previous thread
  • New Chat button for starting fresh

Why It Works

  • Continuity auditors expect from any modern tool
  • Visible persistence builds confidence in detailed queries
Conversation history sidebar with recent chats and New Chat button
Fig 19Conversation history panel — recent chats are always accessible, giving auditors continuity across sessions and confidence that their work persists.
04
Document Intelligence — working with your files.
AccessibilityProgressive disclosure

The Problem

  • Flat, unsearchable list of every file
  • No filtering, no pagination, no feedback on selection

The Decision

  • Two tabs: "Use existing files" and "Upload a new file"
  • Existing files: progressive disclosure — audit name first, then paginated list
  • Upload: drag-and-drop with clear constraints

Why It Works

  • Separates two intents — existing files vs. new upload
  • Optimizes each path independently
05
Smarter data tables — focused review.

The Problem

  • Single scrolling list — no filter, search, or pagination
  • Focused analysis meant scanning every row manually

The Decision

  • Search, pagination, and clear column headers
  • Inline actions (Generate Summary) per row
  • "Continue to Q&A" button once files are selected

Why It Works

Respects the user's time — no one should scroll past 60 irrelevant files to find one document.

Smarter data tables with search, pagination, and inline actions
Fig 20Redesigned data tables with search, pagination, and inline Generate Summary actions — turning a wall of files into a focused review experience.
06
Visibility of system status — no more guessing.
AccessibilityMake state perceivable

The Problem

  • Zero feedback during processing
  • Users unsure if the system was working, broken, or finished

The Decision

  • Full-screen loading state with progress indicator
  • Background dimmed to block interaction mid-process
  • Toast notification confirms completion

Why It Works

  • Nielsen's #1 heuristic — silence reads as failure past 30s
  • Visible progress maintains trust, prevents duplicate submissions
Visibility of system status — loading state with progress bar and toast notification
Fig 21System status visibility — loading state with progress bar over a dimmed background, toast confirming generation. Replaces the beta's silent processing and maintains trust during longer operations.
07
Persona selection — tailored AI perspectives.

The Problem

  • Stakeholders (audit managers, compliance, CROs, IT risk) need different analysis from the same data
  • One-size-fits-all AI response served no one well

The Decision

  • Persona modal with 4 personas, clear focus-area descriptions
  • Selection changes response style + analytical lens mid-conversation
  • Persona prefix ("AM:") confirms the active lens

Why It Works

A UX solution to a prompt-engineering problem — the interface handles the context-switching instead of the user.

Persona selection modal with four analytical lenses
Fig 22Persona selection — users choose an analytical lens (Audit Manager, Compliance Officer, CRO, IT/Cyber Risk) that tailors GenIAus's response style. The persona prefix in the input field confirms which lens is active.
08
Prompt Library — institutional knowledge at your fingertips.
AccessibilityReduce memory burden

The Problem

  • Advanced users wanted proven, reusable prompts
  • Org wanted to standardize AI use across teams

The Decision

  • Searchable Prompt Collection modal, two tabs
  • "My Prompts" (personal, 7/25 used) and "System Prompts" (org-wide, 25 available)
  • Domain-specific, ready to run

Why It Works

Bridges individual productivity and org standardization: users get a head start, the org gets consistency.

Prompt library — My Prompts and System Prompts tabs with searchable list
Fig 23Prompt library — personal and system-level prompts, searchable, domain-specific. Bridges individual productivity with organizational consistency.

Accessibility wasn't a pass at the end. It was the design constraint.

In a regulated audit workflow, accessibility and transparency turned out to be the same problem wearing different names — telling every auditor what the system can do, what state it is in, and what happens next.

WCAG used as a design constraint, not a compliance pass run against a finished interface.

Accessibility was a practical part of making GenIAus trustworthy, not a separate layer applied after the redesign. Auditors work where ambiguity is risk, so the decisions that made the AI legible were the same ones that made it operable: descriptive labels in place of vague feature names, structured Quickstart options and prompt suggestions to cut cognitive load, and visible processing states so nobody had to guess whether the system was working.

WCAG 2.2 organises accessible design around four principles — content should be perceivable, operable, understandable and robust. Those four describe what a cautious auditor needs from an AI system almost exactly, which is why the criteria worked as a design filter rather than a checklist. Every barrier I could name mapped to a pattern the product already needed.

Accessibility as a system rule The criteria, and what they changed

Contrast, and never colour alone

Text, input labels, component boundaries and selected states were designed for readable contrast, and state was carried by label, icon and structure rather than colour. AA asks for 4.5:1 on body text, 3:1 on large text, and 3:1 on the visual information needed to identify a control and the state it is in.

1.4.31.4.11

Keyboard reach, and a focus state you can see

Quickstart cards, chat history, document tables, persona selection and the prompt-library modal are all reachable by keyboard with a visible focus state. Focus cannot disappear behind the processing overlay or a modal layer — orientation survives the deepest part of the workflow, which is exactly where losing it costs the most.

2.1.12.4.72.4.11

Explicit actions, targets you can hit

“Find in Files” became a label that says what it does. Prompt suggestions, persona descriptions and upload constraints removed the assumption of prior AI knowledge. Interactive controls were sized and spaced against the 24×24 CSS pixel minimum WCAG 2.2 sets at AA, subject to its exceptions.

2.5.83.3.2

Status the system announces, not status you infer

Full-screen processing feedback, progress indicators and completion toasts made AI activity visible. The implementation note that travels with them: these need to be exposed as programmatically determinable status messages, so a screen-reader user gets the same update without focus having to move to receive it.

4.1.3

Treating WCAG as a constraint rather than a checklist meant accessibility reinforced the thing the product was already trying to earn — trust, control, and an equal footing for every auditor using it.

Not a handoff. A shared problem.

Redesigning a live product with 8 developers required a different kind of collaboration. Here's what worked.

Findings framed as shared problems got buy-in; prescriptions got resistance.

From a beta auditors avoided to a tool they requested.

−40%Task completion time
across the redesigned flows
+35%Adoption rate
after iteration ship
+84%Time spent on platform
(deeper engagement, not friction)

These patterns became EY's broader AI product design standards.

What I'd do differently. What it taught me.

Transparency isn't a feature in regulated AI — it's the foundation everything else stands on.

What I'd do differently

More upfront co-design with auditors. Workflow nuances only surface when you watch someone do the real work over days, not hours.

Earlier investment in a design system. Defining patterns as I went cost consistency I'd have kept with an upfront design-system sprint.

What it taught me

Transparency is structural, not optional. In regulated domains, every decision has to survive a regulator's questions.

Context retention improves efficiency. History, file persistence, and persona memory aren't features — they're respect for the user's time.

The best AI interface is invisible. When GenIAus works, auditors think about the audit, not the interface — the trust promised up top, delivered.

Next Case Study 02 — Genie Embedding AI into the workflow — not beside it →

Have a project in mind?

I'm open to senior product design roles, advisory work, and selective collaborations. Whether you have a defined brief or a fuzzy problem space, let's talk it through.

damleaalvee@gmail.com
Let's
create.