Client reports increased anxiety following a stressful work situation. Sleep disrupted.
Subjective
Engaged and motivated. Discussed coping strategies and treatment goals. Applied CBT techniques.
Assessment
Generating from session transcript…
Eleos
Session active
Recording00:23:47
AI Draft, In Progress
End Session
Before AI can write the note, it needs to hear the session.
Embedded Audio is the in-app recording layer that makes AI-generated clinical notes possible. Without clean, reliable audio capture, there's no transcript, and without a transcript, there's no note. I designed the base web product and error state system that launched in alpha in July 2024, and became one of the most-used features in the company.
Role
Lead Product Designer
Timeline
2024 (alpha to launch)
Platform
Web (in-app)
Collaborators
Product, Engineering, Clinical Success
NDA note
This work is covered by an NDA, so I don't share screenshots or internal artifacts. Embedded Audio is a live product used by customers. This case study focuses on the thinking behind it: the research, directions explored, and the key decisions and tradeoffs that shaped what shipped. I'm happy to walk through the product and design work in more detail in conversation.
Context
Eleos's core value proposition is AI-generated clinical notes. But AI doesn't work without audio.
Behavioral health clinicians spend an average of 2–4 hours per day writing documentation. The core Eleos product addresses this with AI-generated clinical notes. But the AI pipeline has a hard dependency: it needs a transcript. And a transcript needs clean, complete audio of the session.
Early on, Eleos used two approaches to capture audio: a physical device called Sensi (a dedicated recording hardware piece), and telehealth bots, virtual attendees that joined video calls to capture the session. Both solved the core capture problem. But both came with friction that limited scale.
The question became: what if the recording happened inside Eleos itself, native to the clinician's existing workflow, with no separate hardware and no bot visible to the client?
The Problem
Two prior solutions, and why neither scaled the way we needed.
Prior solution 1
Sensi (Physical Device)
A dedicated hardware recorder placed in the room during in-person sessions. Clean audio, no software complexity.
Hardware dependency. Didn't work for telehealth. Setup complexity limited adoption.
Prior solution 2
Telehealth Bot
A virtual attendee joins the video call and records the session, works for remote sessions without additional hardware.
Visible to clients. Orgs reported discomfort, "will my clients trust a bot in the room?" Limited to telehealth only.
Embedded Audio
In-App Recording
Record directly from the browser. Works in-person and for telehealth. No hardware, no bot, no separate app, native to the clinician's existing workflow.
Single point of capture. Seamless session start. No onboarding overhead for orgs.
The silent failure problem: If audio capture breaks and the clinician doesn't know, they finish a 60-minute session and discover their note has nothing in it. In therapy, you can't recreate a session from memory an hour later.
The design brief: make recording as invisible as possible when it's working, and impossible to miss when it's not.
What I Built
Base product and error states for web. Mobile was explored separately, and deferred.
Embedded Audio touched several surfaces, and not all of them were mine to design. Here's how scope was divided:
My work
Base product, web
Session start / stop flow, recording status indicator, session management, and the full error state system. Designed for the web product where the majority of clinical documentation happens.
Separate track
Mobile exploration
A parallel mobile version was explored by another designer. Mobile-specific interaction patterns and form factor adaptations were scoped separately. The mobile version was ultimately not developed beyond exploration.
The initial scope was web only. Clinical workflow at launch organizations was overwhelmingly desktop and browser-based. Shipping a focused web experience first let us learn from real sessions before adding platform complexity.
The capture session flow, configuration before recording starts, active recording with live timer:
Step 1 · Configure
Capture Session
Let's get started with your session:
Session Type
Individual▾
Setting
In Person▾
Note Type
DMP Note▾
Audio Input
Macbook Pro Speakers▾
Client
Casey Jones▾
+
Session Language
English▾
Capture Session
→
Step 2 · Recording
Casey Jones
Recording...
24:07
End Session
The two-screen capture flow: configure session parameters and verify audio input, then record. No hardware. No bot joining the call. Just start.
The base product, Eleos sidebar recording while the clinician works in their EHR:
app.clearchart.io/notes
ClearChartProgress NoteTreatment PlanDocuments
Jane DoeDOB: 01/01/1980 · MRN: 0000001 · Individual Therapy · 50 min
Chief Complaint
Subjective
eleosSession active
Recording00:23:47
AI Draft, In Progress
Subjective
The Eleos sidebar sits alongside the clinician's EHR. The recording runs in the background, no hardware, no bot. Just press start and focus on the client.
Error States
Audio failure during a therapy session is uniquely high-stakes. Every error state needed a different answer.
The core design challenge was timing: the clinician is in the middle of a session with a patient. We can't interrupt the session for every technical event, but we also can't silently fail. The question for each failure mode was: does the clinician need to act right now, or do they just need to know?
That framing drove the taxonomy. Five distinct failure modes. Two categories:
Requires immediate action, audio can't continue without it
Audio Disconnected and Network Error both mean recording has stopped or is at serious risk. The clinician needs to do something right now, reconnect their mic, switch networks, or the session goes unrecorded. These states are persistent and prominent. They don't auto-dismiss.
Design principle: can't miss it, can't dismiss it without acting
Informational, recording continues, but the clinician should know
Unstable Connection and Session Ending are different. Unstable Connection means quality is degraded but the recording is still happening. Session Ending is a 5-minute advance warning before the session auto-terminates, proactive, not reactive. These states can auto-dismiss or be acknowledged. They don't require stopping the session.
Design principle: visible but not blocking, acknowledge and continue
Conflict resolution, a separate, specific case
Session in Progress is its own category: the clinician tries to start a new recording but another session is already active. This requires an explicit decision, end the previous session, before proceeding. It's the only state with a destructive action button ("End Session"), and the only state where we show that button prominently rather than as a secondary option.
Design principle: make the required action obvious, no ambiguity about what happens next
How an alert surfaces inside the sidebar during a live session:
app.clearchart.io/notes
ClearChartProgress NoteTreatment PlanDocuments
Jane DoeDOB: 01/01/1980 · MRN: 0000001 · Individual Therapy · 50 min
Chief Complaint
Subjective
eleosSession active
Audio Disconnected
We can't hear you. Please check that your microphone is connected and working properly.
AI Draft, Paused
Subjective
The alert surfaces inside the Eleos sidebar, the clinician's EHR stays untouched. Persistent, can't be dismissed without fixing the issue, since recording has stopped.
All five states in the system:
Action required
Audio Disconnected
We can't hear you. Please check that your microphone is connected and working properly.
Action required
Network Error
You are experiencing network issues. Please check your network connection or switch to a different network.
Informational
Unstable Connection
Your Internet connection is unstable.
Informational
Session Ending
We are close to the end of your session. The session will automatically end in 5 minutes.
Conflict resolution
Session in Progress
It looks like you may have another session currently in progress. Please end that session prior to starting a new one.
End
Session
All five error states, designed from Figma. Copy is taken verbatim from the shipped design.
Copy matters here. Each error state is specific about what's wrong and what to do: "We can't hear you. Check your microphone" rather than "An error occurred." That specificity is possible because engineering could surface the exact failure mode. We knew which state triggered. The copy reflects that.
Impact
Alpha launched July 2024. Within months, Clinical Success was calling it one of the most critical features in the company.
July 9, 2024
Alpha launch. A clinician at a regional behavioral health org became the first clinician to use Embedded Audio in a live session.
August 2024
First full org rollout. A behavioral health organization became the first to deploy at scale. User feedback started coming in through CS channels.
Fall 2024
Early outcomes. Clinicians reporting 70% of note work done during the session, a meaningful shift in how notes get written. Newer clinicians gaining confidence from having their sessions captured.
Ongoing
National deployment. Multiple organizations across the country using Embedded Audio as the primary capture method for AI notes.
"I love it so much. This is saving us so much time!"
Clinician · a behavioral health organization
"It helps me focus more on clients during sessions rather than trying to remember everything I need to write later."
Clinician · a behavioral health organization
"EA Sessions seem to be pretty consistently successful and we're getting good feedback."
Clinical Success · Internal review
"Embedded Audio is one of the most critical features in the company right now."
Clinical Success, Eleos
"No more third-party notetakers. Better stability, better audio quality, easier workflow."
Engineering Lead, Eleos · on EA Beta launch
The clinician impact goes beyond time saved. Several orgs noted that newer clinicians gained confidence knowing their sessions were captured; they could focus on the client rather than mentally pre-writing the note. The recording was reliable enough that people actually trusted it.
Reflection
What I'd revisit, and what shipping fast got right.
Session conflict detection could be earlier
The "Session in Progress" state surfaces when a clinician tries to start a new recording while another is active. Ideally, we'd surface that earlier, before the clinician initiates, rather than interrupting them in the act. A more proactive session state check on app load could have eliminated a frustrating moment entirely.
Error state design was a collaboration, not just a design problem
The taxonomy only exists because engineering could tell me what states were actually detectable. Not every possible failure mode was surfaceable with certainty. Working closely with the team to understand what was knowable shaped every decision: which states got specific copy, which had to stay vague, and whether a state needed user action or just awareness.