Okay, Let's talk about design :)
Health care systems aren't failing patients for lack of data. They're failing because no one connected it. Innovaccer is fixing what fragmentation broke with the help of AI.
Project Overview
This project highlights how design helped care managers save approximately 70% of the time manually analyzing risk signals of a client and increasing top of the funnel.
What
Business Impact
<60 secs
80%
Understanding the overall
client risk
-90%
Fewer "surprise" status discoveries mid-call
+78%
Action intent rate on AI insights
Product Design, end to end ownership, strategy & user testing
My Role
1 UX researcher, 2 product managers & AI/ML team & SHIE case managers
Collaboration with
Showing a quick overview of the client !
BACKSTORY
Manually verifying clinical, social, behavioral, and substance abuse data across 50+ daily cases on an average costs case managers 120 hours annually, time spent reading records & not providing care.
Here along with marketing and sales campaigns, design can help optimise these numbers. This particular feature played an instrumental role in saving case manager's time understanding the cross-sectional client risk and further optimizing the overview funnel.
Care coordinators were drowning in data
PROBLEM STATEMENT
30% of the total time in a day was wasted in understanding how vulnerable is the client. The reason varies in different steps of the process and is being discussed in the respective user flows mentioned below. However, some of the generic problems are -
Manual effort of skimming patient info
Speed
Case managers spent 20-25 minutes per client building a mental risk image that should have taken under a minute leaving only 4-5 minutes for the actual care interaction considering a heavy caseload of about 50+ consumers per day.
DISCOVERY METHODS
The data told us what users did. The interviews told us why they suffered.
Deep learning of real-time metrics
& hypothesis
Quantitative Data
Gathering user perspectives &
pain points
User Interviews
I mapped all this into an Experience Failure Matrix, looking at where time pressure + low context = UX breakdown.
1
Reduce manual effort of understanding the client risk
2
Faster context of client's current situation
3
Seamless offline-first experience for usability, accessibility, and continuity.
4
Organic adoption to the overview page of the client
What did we want to achieve for the users ?
TARGET METRICS
PRESENTED BY
Aman Verma - Sr. Product Designer at
Care got faster when I stopped asking
managers to guess
AI decides the risk. You just act.
THE SOLUTION
Manual effort of skimming info
20 mins
Form a mental picture of risk
Old
By AI
60 secs
Risk signals right upfront
New
AI-generated insights are directional, not definitive. The real win is eliminating the manual effort of piecing together a consumer's risk picture from scattered data, so case managers spend less time orienting and more time acting.
But, why do we need a separate tab for a quick
overview about the client ?
Reviewing a client's clinical history, justice involvement, substance use, and behavioral health data separately isn't just slow, it's a design failure. Innovaccer's AI processes the full historical record in one pass, connecting risk signals across domains so the case manager never has to.
How to avoid giving proxy or fake attendance ?
LIVE CLASS ATTENDANCE MARKING FUNNEL OPTIMISATION
2/3
Mark attendance
By Image
Manual
Submit
or
PRIMARY
The Problem Statement
Many times between the student roll calls, a few student helps other students get proxy either by impersonating their voice or by requesting the person who's marking attendance on teacher's behalf.
Voice impersonation
Invalid attendance report
Student roll calls
Existing Flow
New Flow
But, how do we flag invalid students ?
A lot of times, students from another class come and sit in their friend’s class. Those are the
students which the system filters and informs the teacher about them.
Arivoo's AI
Matches live frames with database images
Flags invalid students
What after we went live after 28 days…
In case of similar faces found after uploading class photographs, one student was identified as another student from a different photograph identifying the latter student as present (actually absent).
Wrong attendance
Invalid report
Similar faces
Give Present/Absent
Submit
Fake Attendance
PROXY
By Image
Manual
Facial Recognition
Invalid Students Filtered
or
From inception to execution
VISION
The vision was to learn from quick experimentation and rapid testing with users to offer the best digital attendance marking experience to them.
Rapid Testing
Quick Optimising
TESTING METHODS
How did I validate the solution
I arranged workshops with the customers who were engaged with Innovaccer and conducted a task based research to deeply understand how they are interacting with the solution.
Wizard of Oz Testing
Field Pilot
Interaction patterns study
Who should care ?
THE USERS
The overloaded coordinator
The case supervisor
Our primary target audience has the similar needs as this persona -
Hi, I'm Eliza
A care coordinator at Sacramento County Department of Health Services. I manage a caseload of high-risk clients, connecting them to clinical, social, and behavioral health resources across every stage of care.
Why speed ?
Care coordinators don't have time to be detectives. Every minute spent finding risk is a minute stolen directly from delivering care.
I want
Understand the client's current risk before doing anything else
Decide what action to take next
Document and justify the care decision
Current Care coordinator workflow
Opens SHIE
Searches client
Clinical + Behavioral + …
Provide care
How to reduce decision fatigue understanding key
cross-sector risk signals ?
1/3
We introduced a brand new overview page designed for only for one job - understand the client's current risk before touching any individual data domain.
New - 'Overview tab' for a quick risk signal (Failed experiment)
What if a care coordinator gets a primary entry point that surfaces a client's key status flags, a 12-month encounter history, and active programs, all above the fold, before any deep navigation begins.
Key status flags
Last 12 month encounter history
New Entry Point
Old
How to reduce drop off rate from 31% to 15% ?
To reduce the drop off rate, we optimised the user flow in following 3 layers:
1
Adding a dedicated benefits page
2
Defining strict image acceptance criterion
System Efficiency
3
Alert for similar faces
Risk Management
Winning Trust
Funnel 1 - Winning Trust
ADDING A DEDICATED BENEFITS PAGE
Solving for trust
Distrust comes from unfamiliar things. Here user’s concern would be ‘What if the new attendance marking process is not accurate enough, do I again have to recheck manually ?’
How did we address this issue ?
Transparency
Clear Communication
Validation
What is the new process
How does it work & benefits
Social proofs for inefficiency
Frustration
USER RESEARCH
The existing ‘How does it work’ section was inferred as a distracting estate for the teachers. Also, teachers expected to see student list right after clicking on mark attendance instead of forcing them to upload a class photograph.
PROPOSED OFFERING
To enhance clarity and focus, we introduced a dedicated benefits page that highlights the value of Quick Attendance. This allowed the main experience to stay streamlined, keeping student names front and center, and making it easy for teachers to quickly mark attendance with a class photo.
1st Revamp (discussed above)
/ Roll Calls
/ Forcing user to upload class photo
The core USP : Mark attendance of the large student group within seconds
What
Dedicate only upto 2% of the total
class time marking student attendance
Outcome
A self explanatory process for quick attendance : Capture, upload and wait for a few secs
How ?
Mentioning the benefits to make
teacher realise how efficient is this way of marking student attendance
Why ?
Winning trust of the teachers by showing social proofs & making them realise the actual need for quick attendance.
Need
/ Asking to upload photo for faster process
Bottom sheets - Benefit cards
SECONDARY LINK TO GO & MARK ATTENDANCE
Credibility
/ Keeping the core experience intact
/ Winning Trust
Funnel 2 - System Efficiency
STRICT CRITERION FOR IMAGE UPLOAD
Low/No Internet
Poor Image Quality
Improper Lighting
Obstructed Faces
Image rejection use case
If the system doesn’t accept the image, user still has an option to skip and give manual attendance.
Also for a higher adoption rate, I made the benefit clear and control visible. Users preferred helpfulness rather than ambiguity.
EDGE CASE 3
EDGE CASE 4
EDGE CASE 5
EDGE CASE 6
Not accepted
Accepted
Image acceptance use case
Once an image gets accepted by the system, the focus shifts to ‘proceed’ keeping in mind that
user might want to add more images as well.
But, how does the images get processed ?
Schedule a meeting and I'll answer all your questions.
aman.vermaproduct@gmail.com
—
EDGE CASE 1
EDGE CASE 2
Funnel 3 - Alert for Similar Faces
RISK MANAGEMENT
Wrong attendance
Trust issues
Invalid report
Opportunity
The similar faces will be flagged to the teacher before providing the recognised student list.
The teacher will manually confirm if the student is actually present or not.
1/3 - Upload
EXISTING
2/3 - Similar Faces Found
NEW
3/3 - Students Recognised
EXISTING
Challenges in User Experience
The teacher has an option to upload multiple images. A student was recognised as another student due to a different angle, lighting, or any other external factor in the other image identifying the latter student as present which was wrong,
Happy Flow - Quick Attendance by AI
NEW HASSLE-FREE EXPERIENCE
< 1 min
Comprehensive error handling
Reduced no. of clicks
Scroll Left to see the happy flow
TOUCH POINT
MANUAL/AUTOMATIC
1/6
Start
CONTROL
MANUAL/AUTOMATIC
2/6
Captures
CONTROL
MANUAL/AUTOMATIC
EDGE CASE W/ PROPER CONTEXT
IMAGE REJECTED
3/6
Captures again
EDGE CASE W/ PROPER CONTEXT
IMAGE REJECTED
EDGE CASE W/ PROPER CONTEXT
IMAGE REJECTED
4/6
Proceeds
ANTICIPATION
IMAGE REJECTED
…
Waits
MOMENT OF TRUTH
IMAGE REJECTED
5/6 (Less Chance)
Manual Check
EDGE CASE W/ PROPER INFORMATION
SIMILAR FACES
EDGE CASE W/ PROPER INFORMATION
SIMILAR FACES
FRICTION FLOW
AI OUTCOMES ARE NEVER 100% ACCURATE
6/6
Submits
TOUCH POINT 2
Nth TIME EXPERIENCE
3/3
How to make sure the system works even in low
or no internet access ?
GRACEFUL DEGRADATION & FALLBACK FLOWS
< 500kbps
Low/No internet
Challenges in User Experience
A lot of times big universities are set in the outskirts of a city. Even after having a wifi access, the internet is many times unstable specially within the university. Perceived abandonment lead to restarting the app or switching back to manual way of marking student attendance ultimately wasting time.
Opportunity
Teachers will click photographs and later will be able to upload from the gallery whenever the internet is available even post class.
Focus is on usability, accessibility and continuity.
Users care more about how stable the experience feels rather than how accurate the signal actually is.
Is the quick attendance funnel really optimised ?
These are the max no. of taps user has to do considering the edge cases as well.
6 Taps
The complete live class attendance marking funnel gave results within a minute unlike the manual process which took around 6 mins depending upon the student strength.
Where did we win ?
Time
Dashboard - Masthead States
COMPLETED
LIVE
Online
UPCOMING
COMPLETED
LIVE
UPCOMING
Offline w/ Location
Overall Business Impact
Latest numbers since launch (14th April, 2025) : Design & Tech
89%
Attendance via photo
upload initiation rate
67%
<1 min
TOT
(Time on Task)
85.8%
-61.2%
Drop offs while marking attendance via photo upload
+8%
More time for student
interaction
0%
Proxy
(fake attendance)
PRESENTED BY
Aman Verma - Sr. Product Designer at Innovaccer (Series F, $675M at $3.4B valuation)
Ex -
(10M+ Userbase)
Default
Low Internet ( <500kbps)
Live Class
Post Photo Upload
