2022-2025 / SaaS / Health / Web - Tablet

Faster triage and smarter replies for front desk staff

Faster triage and smarter replies for front desk staff

3x faster

to find and send the right template

67% → 34%

drop in hesitation before sending

My Role

Senior Product Designer, Practice Workflows:

Designed quick inbox actions and LLM-assisted messaging, from early concepts through to rollout. Owned UX direction and worked with Product, Engineering, and Machine Learning teams throughout.

The challenge

The inbox and message thread were minimal by design. But as the product grew and customers faced more patient messages, every small interaction and delay was magnified. Workflows became inefficient and frustrating.

My goal was to find these moments and to join the dots, making repetitive workflows smoother and, in some cases, eliminating portions of them entirely to lay the foundation for a more powerful workspace.

Refining the problem

We trawled the data we had on staff work flows from our Looker analytics to customer feedback both existing and gathered as part of our primary and secondary research process (customer interviews, surveys, benchmarking) to find the main pain points facing users.

The results varied, but we were able to identify a cluster of related issues falling into two main subsets:

  • Friction transitioning between inbox and patient thread

  • Time taken to understand and respond to a patient message

Later, as directions solidified I ran a series of concept validation measures to put early ideas in front of customers that gauged appetite and gathered initial feedback.

What to tackle

1

Process of needing to open a message thread to perform any action was slow and out of date vs analogous apps

2

Some workflows in the thread were cumbersome and strong candidates for the emerging AI layer we were experimenting with

Solution 1

Solution 1

Speedier triage

one interaction at a time

As customer needs grew, the simple inbox became limiting. Staff needed faster ways to act on messages.

Problem

Even simple actions meant opening every message. Staff had to click into each thread to assign, archive or reply, even when the preview told them what to do. With hundreds of messages a day, those extra clicks added up fast.

Hypothesis

If the most common actions sat directly in the inbox list, staff could clear routine messages without opening them. Opening a thread would become a signal that something needed attention, not a step every message demanded.

We'd know it was working when [signal, e.g. less time spent sorting per shift, with no rise in misrouted messages].

New inbox

Old inbox

Message counter

Message counter

Message counter

Flags multiple unread messages as a quick urgency cue.

Flags multiple unread messages as a quick urgency cue.

Flags multiple unread messages as a quick urgency cue.

File indicator

File indicator

File indicator

Surfaces attachments that might affect how a message gets triaged.

Surfaces attachments that might affect how a message gets triaged.

Surfaces attachments that might affect how a message gets triaged.

Content preview text

Content preview text

Content preview text

A line of content to give better context at a glance.

A line of content to give better context at a glance.

A line of content to give better context at a glance.

Voicemail preview widget

Voicemail preview widget

Voicemail preview widget

Makes voicemails obvious in the list and playable directly in the inbox.

Makes voicemails obvious in the list and playable directly in the inbox.

Makes voicemails obvious in the list and playable directly in the inbox.

Moved to backlog

Moved to backlog

What this unlocked

These new inbox elements set the platform for quick actions, so staff could act on a message without waiting for the thread to load.

So we took a common pattern any user would be familiar with - a hover state that reveals more actions. In our first version, these were the ability to triage a message to an inbox or to close it.

This covered the vast majority of first actions our target persona of Front desk staff were responsible for.

Using these additional inbox elements, we set the platform for introducing quick actions to enable staff save a click and load time of the conversation thread view before performing an action.


So we took a common pattern any user would be familiar with - a hover state that reveals more actions. In our first version these were the ability to triage a message to an inbox or to close it.

This covered the vast majority of first actions our target persona of Front desk staff were responsible for.

These new inbox elements set the platform for quick actions, so staff could act on a message without waiting for the thread to load.

So we took a common pattern any user would be familiar with - a hover state that reveals more actions. In our first version, these were the ability to triage a message to an inbox or to close it.

This covered the vast majority of first actions our target persona of Front desk staff were responsible for.

Better context + Actions upfront

Faster, accurate message triage

Solution 2

Utilising AI

My team worked closely with a Google AI engineer, on a one-year secondment, to explore where LLMs could help in Klara workflows.

My team worked closely with a Google AI engineer, on a one-year secondment, to explore where LLMs could help in Klara workflows.

After ideation and feasibility sessions, we decided to focus on two key elements first:

After ideation and feasibility sessions, we decided to focus on two key elements first:

Response template suggestion

Response template suggestion
Problem

Problem

Practices had built up large template libraries, but finding the right one was slow. Staff scrolled or searched by name, then edited the text to fit the patient. Many skipped templates altogether and typed replies from scratch.

Hypothesis

Hypothesis

If the inbox suggested the most relevant template based on the message, staff would use templates more and write from scratch less. Keeping suggestions editable would let them send with confidence while still making the final call.

Voicemail
transcription

Voicemail
transcription
Problem

Problem

Voicemails had to be listened to in full, often more than once. Staff noted names, dates of birth and requests by hand, then decided what to do next. A short message could take several minutes to process.

Hypothesis

Hypothesis

If voicemails arrived as text with key details and suggested actions pulled out, staff could handle them like any other message. They would scan instead of listen, and only replay the audio when something was unclear.

Response template
suggestion

Existing interaction

Template selection required digesting the patient response then navigating often cluttered and poorly maintained template menu housed in the composer bar.

Our solution

Our solution

Staff were losing minutes on every reply, scrolling through templates to find the right one. Now the LLM reads the thread and puts the most relevant templates right where they're typing. They can accept one, tweak it, or bin it and carry on.

Training the model
Training the model

Before suggesting message templates, we gathered feedback from practice staff to sharpen the model’s accuracy.

I designed a feedback module so staff could confirm our analysis or suggest a better category. Once predictions were solid, we piloted the suggestion module with 10 customers and expanded as accuracy and trust grew.

Below is an assumptions map of the logic I put together with engineers to work out how we could gather continuous feedback on the suggestions we serve up.

Before suggesting message templates, we gathered feedback from practice staff to sharpen the model’s accuracy.

I designed a feedback module so staff could confirm our analysis or suggest a better category. Once predictions were solid, we piloted the suggestion module with 10 customers and expanded as accuracy and trust grew.

Below is an assumptions map of the logic I put together with engineers to work out how we could gather continuous feedback on the suggestions we serve up.

Before suggesting message templates, we gathered feedback from practice staff to sharpen the model’s accuracy.

I designed a feedback module so staff could confirm our analysis or suggest a better category. Once predictions were solid, we piloted the suggestion module with 10 customers and expanded as accuracy and trust grew.

Below is an assumptions map of the logic I put together with engineers to work out how we could gather continuous feedback on the suggestions we serve up.

Voicemail actions

Klara's voicemail feature let staff play the audio of a patient's call. But listening to audio in a busy clinic caused delays and broke the flow of their text-based work.

We used a custom LLM, trained on third-party medical data and our own growing data set, to transcribe voicemails accurately. That solved the foundational issues and opened up new features.

Jumping into the message

Staff told us they replayed messages again and again to catch names, dates, and medications. Transcriptions broke audio into clickable words, opening up new features.

Transcription sped up info processing, but staff still needed more trust. We let them hover or click a word to instantly jump to that spot in the audio in order to double check that detail.

What came next

The next step was layering meaning on top of the transcript: keywords, a short summary and quick links to relevant actions, so staff could act without pressing play. This was planned for later iterations, once the first use cases were proven.

Outcomes

Quick actions and AI template suggestions both rolled out across Klara's customers, and front desk staff picked them up fast.

94%

adopted quick actions in 30 days
Triage without opening the message

3x faster

to find and send the right template
78% of staff adopted AI suggestions

67% → 34%

drop in hesitation before sending
Staff replied with more confidence

What we learnt

Quick actions eased the jump between the inbox and the thread, but didn't fix it. Staff picked them up fast, but without more context they often still opened the message. A multi-pane view would have solved that. With a small team and an acquisition coming, we never found the time. Looking back, I'd have tested a rough version with customers early, rather than waiting for room to do it properly.

  • Quick actions eased the jump between the inbox and the thread, but didn't fix it. Staff picked them up fast, but without more context they often still opened the message. A multi-pane view would have solved that. With a small team and an acquisition coming, we never found the time. Looking back, I'd have tested a rough version with customers early, rather than waiting for room to do it properly.

We built the AI features before today's plug-and-play models were available. They'd make a lot of this easier now, though HIPAA would probably still have pushed us towards something bespoke.

  • We built the AI features before today's plug-and-play models were available. They'd make a lot of this easier now, though HIPAA would probably still have pushed us towards something bespoke.

✧