Automate the obvious. Keep humans in control.

Every field was entered by hand, then checked by someone else.

Finance and operations teams received invoices, contracts and purchase orders by email or shared folders. They opened each document, read it, entered the data, and then someone else checked it. After that, it went for approval.

What this caused

• Slow process — one document moved through several people and tools.
• Repeated work — the same data was read twice: once to enter it and once to check it.
• Scattered status — updates were spread across emails and folders.
• Poor visibility — managers could not easily see what was pending.
• Late errors — mistakes were often found during review or approval.

I owned the data extraction and review workflows.

I was the designer for the extraction and review workflows. I worked on research, user flows, UI design and prototypes. I worked closely with the PM, ML lead and engineering team.

Another UX designer worked on other parts of the product, such as audit trail, questionnaires and settings. We shared design patterns to keep the product consistent.

What I owned

• Data extraction and review — how a document moved from upload to approval.
• Confidence scores — how users understood AI confidence.
• Review and validation — how users checked and corrected the data.
• Task and handoff — how work moved between people.
• Dashboards — what each role needed to see.
• UI and design system — components, patterns and prototypes.

What made it hard

• The AI could make mistakes, so users needed a way to check the results.
• The product was new, so users were not familiar with it.
• The confidence score came from the ML model, so I worked with the ML lead to make it easy to use.
• We wanted to make the process faster while keeping people in control.

What I wanted to understand.

Before designing, I wanted to understand how people did the work, where they spent the most time, and where mistakes happened.

What I wanted to understand

• Where do people spend the most time checking?
• Why do they check the same data twice?
• What helps them trust the data?
• What should AI do, and what should people do?

People I spoke to

What I found

Let AI handle the easy cases. Let people handle the decisions.

Users wanted less manual checking, but they still wanted to check important values. I used confidence scores to show where they should focus.

01 — Show confidence only when it helps

Decision

Users did not need to check every field. I highlighted the fields that might be wrong.

Why

This helped reviewers move quickly through correct values and focus on the ones that needed attention.

02 — Keep the document next to the data

Decision

Users could compare a value with the original document without leaving the screen.

Why

When users were unsure, they could check the original document right away.

03 — Send uncertain values to a person

Decision

AI handles the values it is confident about. Uncertain values go to a reviewer.

Why

The goal was to reduce unnecessary checking, not remove people from the process.

What I decided not to do.

• Fully automate everything. Some AI results could still be wrong, so I kept people involved when the AI was unsure.
• Show confidence on every field. Too many scores would add noise, so I showed them where they helped users decide what to check.
• Show only an edit option. Users needed both a quick fix and a way to check the source, so I kept the document visible.
• Use separate screens for each step. I kept extract, check and approve in one place so users could finish a document without moving between screens.

Seven parts, each with a clear purpose.

Reviewers spend most of their time on this screen, so each part has a clear purpose.

1. 1Product railGives access to the dashboard, inbox, review queue and templates. Labels help new users learn the product.
2. 2Question groupsShows what the AI needs to find. It also works as a filter, so reviewers can focus on one group at a time.
3. 3AI assistantExplains why a field was flagged and can suggest a value. It does not make the final decision.
4. 4Document canvasShows the original document. Highlights show which values need attention.
5. 5Confidence popoverShows the confidence score, why the field was flagged, and options to check or fix it.
6. 6Review panelShows review progress, confidence levels and fields that still need a decision.
7. 7Edit and source drawerOpens beside the document. Users can edit the value or check the source while keeping the document visible.

One workflow, different views for different roles.

The same document moves through different roles. Each role needs different information: reviewers need to know what to work on, team leads need to see where work is stuck, and managers need to see changes and errors.

The reviewer view — "What do I need to work on?"

• What is assigned to them
• What needs review
• What is urgent
• What has been waiting too long

The goal: help reviewers find their next task without asking anyone.

It has not shipped yet, so here is how I would measure it.

Because the product has not launched yet, I would track these five measures.

• Time to review one document.
• How many fields users still check by hand.
• How many documents wait for more than a day.
• Documents completed by each reviewer per day.
• How often reviewers change an AI value — a useful sign of how well the AI is working.

What I learned

Start with the user decision.Understand what the person needs to decide before deciding where AI should help.

Show uncertainty only when it helps.A confidence score is useful when it helps the user decide what to do next.

Design for mistakes.AI can be wrong, so the review flow needs to work well when that happens.