BUILT THE PRACTICE

AI Docflow — Workflow Optimization at Scale

An AI-powered document workflow designed to reduce manual data entry, simplify task tracking, and make automation more transparent and trustworthy for finance and operations teams handling large volumes of documents.

Project: AI Document Processing Platform
Role: Lead Product / UX Designer (end-to-end)
Timeline: Started 2023 · Currently in active development Domain: Enterprise SaaS · Finance & Operations

OVERVIEW

Building AI workflows teams can trust.

Finance and operations teams drown in documents — invoices, contracts, purchase orders — and most of the work is still done by hand. The current process is slow, full of errors, and impossible to scale.

We designed an AI-powered workflow that automatically reads and processes documents while keeping humans in control. If the AI is confident, it completes the task automatically. If not, it asks for human review.

This created a system that is faster than manual work and more reliable than fully automated workflows.

Note: This product is in active development. The numbers above are design targets set by benchmarking against the manual workflow it replaces. They will be validated against real users at launch.

CHALLENGE

Manual processes that couldn’t scale.

Finance and operations teams handle hundreds of documents every month — invoices, contracts, purchase orders. Every step was manual: read the document, type the data into the system by hand, email the next person to assign the next step, then chase status updates over Slack and email. As volume grew, the process couldn't scale. Adding more people only added more cost and more places for things to go wrong.

RESEARCH

I started by sitting next to people doing the work, not by drawing screens.

Before designing anything, I spent time understanding how the existing manual process actually broke down — where people got stuck, what they didn't trust, and where the system failed them.

Key Insights That Guided The Design

STRATEGY

Automate the obvious. Keep humans in control.

Full automation would have been faster, but reviewers wouldn't have trusted it. A fully manual queue would have defeated the purpose. So the workflow splits by confidence.

When the model is sure, the document moves on. When it isn't, the task goes to a reviewer with full context — what the AI saw, why it wasn't sure, and what it suggests. Every correction feeds back into the model, so the system gets sharper over time.

Fast where speed is safe. Human where judgement matters. Trust earned through visibility.

DESIGN DECISIONS

The AI handled what it understood. Humans handled the uncertainty.

Four decisions came out of this principle:

THE SYSTEM IN PRACTICE

A five-step workflow that prioritises low-confidence cases and reduces manual work.

Step 1 — Empty state. Before any document is uploaded, the workspace stays calm and focused. There's exactly one thing to do: upload a document. Everything else (chat, navigation, secondary features) stays in the background until it's needed.

Step 2 — Upload. Users can upload one or many documents at once. A progress indicator shows real-time status for each file, so users always know what's happening. No silent failures, no guessing.

Step 3 — Extraction overview. Once processed, the system shows three things up front: which fields were extracted, where confidence is low, and the overall trust level for the document. Users can scan and decide what to do next in seconds.

Step 4 — Confidence-based review. The system highlights low-confidence fields and explains why they were flagged. Users can quickly review the AI's suggestions and either accept or edit them — focusing only on what actually needs attention.

Step 5 — Final review & approval. After reviewing flagged fields, users see the complete extracted data in one organised view. They can verify, edit if needed, and approve the document with confidence — all from one screen.

TRANSPARENCY & TRUST

Can users see, verify, and correct what the AI does?

The system surfaces AI confidence, reasoning, and progress so users can trust automated decisions, step in when needed, and stay in control of every outcome.

How the workflow actually operates.

A good workflow does three things: it moves work efficiently, lets people trust each AI decision, and gives team leads a clear view of the whole. Here's how we designed for all three.

Team oversight

Individual reviewers needed transparency at the field, batch, and document level. Team leads needed something different — one view of what was flowing through the system, what the AI was handling, and where humans were still needed.

The dashboard reads incoming work, sorts it by confidence and urgency, and shows leads where their team's attention is needed. Routine work moves on its own. Exceptions surface clearly. No one chases status in Slack.

BEFORE & IMPACT

What changed in the review experience.

The original review panel didn't tell users why a field was flagged, what to fix, or what to do next. The redesign answers all three.

IMPACT

This product is currently in active development, so the numbers below are design targets, not measured outcomes. They were set by analysing the current manual workflow and benchmarking against existing tools in the space. Once the product launches with its first users, the design will be validated against the manual baseline it's replacing.