Ernst and Young
2025
What if consultants could find the evidence in thousands of documents without opening each one?
EY auditors and consultants review client lease contracts and policies. To find one statement, they opened every document and read it, across many file formats.
Document Explorer, now part of the EY Document Intelligence platform, is an AI search tool that finds the exact sentences that matter. I designed it in 2021, before generative AI went mainstream, which makes it one of my first AI products.

EY consultants that finds the exact sentence that matters, so they spend time judging evidence instead of hunting for it.
Overview
How I designed AI search consultants could trust
Results show the matching sentence and a relevancy score, not a list of documents. Consultants check the AI's answer before they use it with a client. I was the only designer, from research to launch in 5 weeks.
Strategy:
I planned the work in three steps: make it work reliably, make the AI parts usable within technical limits, then make it easy enough for anyone to self-serve.
Impact
Shipped as an MVP, tested with the people using it
I tested working prototypes with 4 consultants in 45-minute sessions over Teams, shared findings with the product owner, and iterated through three designs before launch.
Source: EY Document Intelligence platform.
"We're busiest in audit months. We spend 6 hours a day, and hire interns, just to find evidence."
Interview participant
EY consultant, research interview
"This solution will definitely solve our manual processing time significantly. Loved it."
Participant
EY consultant, usability testing
Problem
Finding one sentence meant reading everything
An EY auditor runs 40 to 50 audits a year. Each one brings 500 to 10,000 lease contracts, from 10 pages for a vehicle to 300 for real estate, and each takes 3 to 4 hours to review.
There was no tool to search them, so every assessment started with hours of reading.

01
Every search meant opening every document
Consultants read through files one by one to find a single statement.
02
Formats didn't match
Policies came in many file types, and each one had to be explored differently.
02
1. Evidence was uneven
Manual searching made assessor reports slow and uneven from one consultant to the next.
Research
What research showed
interviews and a survey with auditors, plus storyboarding and Q&A sessions to learn terms like assessor reports
How AI works
People confirm the AI's work before anything is extracted
"Semantic search finds sentences that mean the same thing, even when the words differ. Question answering pulls the exact answer, like '8 characters'."

Design Decisions
Design Decision 01
Show the sentence, not just the document
The AI could find relevant sentences, but consultants had to stand behind them in front of clients. A list of matching documents would still send them back to reading. Results had to lead with the sentence, its source and a relevancy score, and let consultants confirm or reject it. Getting that review step right took three tries.

Three versions of the review screen: a slider with a pop-over panel, a hover-to-review layout, and the final single page with selected text highlighted.
Trade-off: One tester asked, "Can we check relevancy by ourselves?" A score can make people trust the AI too much, so I made it easy to mark any result right or wrong. Consultants stay the judge, and the AI learns from them.
What it cost: Selecting text is easy to miss the first time, so the final design needed light onboarding.
Craft note: I checked color contrast with Stark and tested full keyboard use (Tab, Shift+Tab, Enter, Space). The yellow and green highlights passed.
02
Reusable query sheets instead of one-off searches
Consultants asked the same questions across many clients, such as finding every statement about multi-factor authentication.
Instead of typing each search again, query sheets let them run a whole set of questions against a new client's documents at once.
A query sheet with a set of assessment questions run against a client's documents, with matching sentences for each question.
Trade-off: Fully editable sheets would have taken longer to build. To keep the MVP small, I limited editing and started with fixed question templates.
If I designed it in 2026
The tools changed. The principles didn't.
In 2021, the AI could find the answer: ask "What is the minimum password length?" and it returned "8 characters". Today it can explain that answer and draft the report. Here is how I would take Document Explorer further.
Answer, then show the proof
A plain question gets a short answer, with every line linked to the sentence it came from.
Explain the confidence
Instead of one score, the AI says why a result matched, and clearly flags when it isn't sure or found nothing.
AI drafts, people decide
The AI drafts the assessor report from approved evidence. The consultant approves each finding before it reaches a client.
Each principle builds on a 2021 decision. Sentence-first results become answers with sources, and right or wrong marking becomes explained confidence and human approval.


