BackSerena Huang
HEALTHTECH · AI WORKFLOW AUTOMATIONUX / PRODUCT DESIGNFRONTEND DEVELOPMENT

AI Document Splitting

Overview

Integrated with email/FAX inboxes and Practice Management Softwares (PMS), Facere Smart Admin automates document handling workflows in medical practices — reading incoming referrals, results, and correspondence, then filing each document to the right location.

Role

Product Manager, UI/UX Designer, Frontend Developer

Period

2 weeks in Aug - Sept 2025

Tools

Figma, Typescript, React

Expertise

Workflow DesignAutomation SystemsHealthcare UX

Problem

Multi-patient documents were breaking the automation

Medical documents get imported the Facere AI system automatically from the users faxes and emails. But as we onboarded larger practices, we hit a recurring edge case: a single incoming document often contained information for multiple patients at once (e.g. a pathology document summary covering several patient's results).

The existing system had no way to handle this. Staff had to pull the document out of Facere entirely, manually split it in an external tool like Adobe Acrobat, and re-upload each piece — undoing the automation the product was supposed to provide.

How might we...

automate the document-splitting process itself, while still giving staff full visibility and control over what the AI decides?

Solution

Automated splitting; Keeping the human in the loop.

We built AI-driven page splitting directly into the document processing pipeline: the system detects when a document contains more than one patient and automatically separates it into individual tasks. My focus was the front end — designing and developing the correction loop that lets staff intuitively review and fix the AI's split decisions before anything gets filed.

Key Features

DecisionWhy
Three-step flow (Patient → Pages → File Contents)Matches how staff naturally reason through a document; keeps each screen to one decision
Persistent per-page Include/Exclude toggle gridAI proposes the split — staff only need to correct mistakes, not rebuild from scratch
Full-page disabled state for excluded pagesStatus readable at a glance, whole document still visible for context
Live page-range drop-down summary ("Page 1–3, Page 5")Lets staff sanity-check the full selection in one glance
Confirmation dialog on submitProvides visibilty into unaccounted pages to prevent user error

Design Process

Discovery

Surfaced through customer success calls: one practice estimated ~20% of incoming documents were multi-patient.

Practices described working around it by exporting documents to Adobe Acrobat — leaving the Facere workflow entirely and still paying for a separate tool to perform a manual process.

Known Design Constraints

Accuracy is non-negotiable. Misrouting a page is a patient-safety issue, not a minor bug — every split had to be inspectable before filing.

Staff are time-poor. The review step had to be fast, or busy staff would rubber-stamp it without actually checking.

Trust had to be earned. Practices had already been burned by the lack of splitting support — the design needed to visibly demonstrate control, not just claim it.

Iteration

Given resource constraints, we ran a two-tier validation strategy: fast internal blitz tests to catch obvious usability gaps early, then real customer conversations to pressure-test the design against actual clinical workflows — before committing engineering time to a build.

V1
Internal blitz test

Expandable panel, open by default

The Include/Exclude control and page-assignment list lived in a panel that defaulted open. I ran blitz usability sessions internally — engineering and product colleagues walked through the Figma prototype using a think-aloud protocol.

The panel was still too easy to overlook if closed and navigated to the next page, and status wasn't visible on the page itself — the only way to tell which pages were included or excluded was reading the numbers in the right panel. Expanding the panel also shrank the document preview, which mattered since staff need to read small print on the page, and added an extra click to a decision that should have been quick.

Finding: Panel easy to overlook; expanding it shrank the document preview and added a click

V2
Customer validation

Always-visible toggle + full-page disabled state

We removed the expand/collapse entirely so the Include/Exclude button was persistent on screen, and added a full-page disabled overlay so status was visible without reading anything.

Taking this to real customers surfaced a sharper concern — large documents (20+ pages) made page-by-page checking too slow, and a page number alone wasn't enough to judge inclusion:

If there's multiple pages, I want to make sure people actually look at the documents.
Practice Manager, QLD
Sometimes we receive documents that are 20+ pages — I know staff wouldn't click through each one.
Practice Manager, NSW
The page number isn't enough — I need to see the contents of each page to know which ones should be included.
Practice Manager, NSW

Finding: Scalability failure: 20+ page documents were still too slow to audit page-by-page

V3
Final design

Numbered instructional sidebar + page summary

We added clear instructional copy (Step 1/2/3) on the side bar so the flow itself couldn't be missed, a page-range dropdown a page-range dropdown summarising every page's status in one place — so managers could audit a 20+ page document at a glance. A confirmation dialog allows user to cross check pages and confirm their selections before submission for additional error prevention.

Roll out & Impact

10+ hours saved in 26 days at one NSW medical practice

We started with an beta rollout, enabling the feature flag for customers who came to us asking for this feature and the ones who helped us refine it.

Feedback was immediate and enthusiastic:

  • "This is a lifesaver – removing an entire step from the traditional process"
  • One practice managed joked that she no longer needed to worry that "my staff are filing away multi-patient documents incorrectly, the doctors will be complaining a lot less."

Later we rolled out to our entire user base; here are the stats for one particular practice in NSW over a 28 day period:

228

Multi-patient Documents

Received within 26 days at the Medical Centre

10.45h

Time Saved

Compared to manual end-to-end multi-patient document allocation processes.

83.3%

Time Reduction

In document processing time.

Time comparison

MethodTime per docTotal (228 docs)
Manual processing~3 min11.4 hours
Facere Smart Admin~15 sec57 minutes

Learnings

Reflections

Designing with AI, keeping the human in the loop

Automated page-splitting reached 95.9% accuracy — strong, but not perfect. What made it viable in a clinical setting wasn't that number; it was the correction loop behind it. A fast, low-friction way for staff to review and fix a split before it's filed is what let us ship an AI feature without waiting for 100% accuracy first.

The manual-check step isn't a gap in the automation — it's what earns staff's trust in it, and it doubles as a feedback signal: 14.3% of corrections were patient-reassignments rather than page reordering, pointing directly at clinician-matching as the next area to improve.

Challenge the ask, find the root problem

The initial customer request was simple: a manual splitting UI inside Facere. Building that would have solved the immediate complaint but left the real problem untouched — splitting documents by hand is still slow.

Pushing past the literal ask to the underlying time cost is what turned a small UI request into an AI-driven feature that meaningfully improved the workflow, not just moved it. Watching practice staff pick it up and seeing it genuinely lighten their workload was the most rewarding part of this project.

Bring end-users in early

Blitz testing with internal colleagues was fast and useful for catching obvious usability gaps, but it couldn't surface needs specific to the clinical context.

The page-summary dropdown — which turned out to be one of the most valued parts of the final design — only emerged once real practice managers were in the room. Involving that broader mix (practice managers, front-desk staff) from ideation, not just validation, would likely have surfaced it sooner.

~ Serena

Next Project

Festival360