Cadence
Combining six tools into one - allowing trainers to take on 3x more agents.
Lead UX Designer
2 Months
Web App
Inefficient workflows made training difficult to scale - impacting caption quality for over 800k people who rely on captions during phone calls. Cadence streamlines training with automation and centralized tools.
Background
What is a Captioning Agent?
Captioning agents listen to live phone calls and repeat what they hear in real time, providing instant captions for people who are hard of hearing.
The Problem
Training captioning agents relied on six disconnected tools, making the process slow, manual, and inefficient.
Trainers
35-40
Manual grading and multiple disconnected tools consumed valuable time and made scaling training difficult.
Agents
850+
Training didn’t mirror real-world scenarios, leaving top performers unchallenged and struggling agents unsupported.
Consumers
800k+
Poor training led to inconsistent caption quality, negatively affecting accessibility for people who are hard of hearing.
Our Team
We were asked to build a simple grading tool
What was originally scoped as a quick solution turned into a deeper look at how training actually worked across tools, teams, and workflows.
Research
Six tools, no structure, and zero visibility
To ensure we built the feature effectively, we spoke with users and stakeholders to understand their needs and expectations. Within days, deeper problems began to surface. We realized that solving these problems required more than a simple test call feature - we needed to reimagine the entire training process.
Disconnected Tools
Training relied on six separate systems for grading, feedback, and progress tracking -slowing down workflows and creating unnecessary complexity.
Unclear Call Selection
With hundreds of test calls and no way to gauge difficulty, trainers struggled to know which calls to assign - often overwhelming or under-challenging agents.
Lack of Performance Visibility
Trainers couldn’t track detailed progression over time, making it hard to deliver personalized training or identify growth opportunities.
Design
Solving for the most time-consuming task first
The first major problem to tackle was the grading process. Trainers and supervisors spent over 7 hours each week marking correct, incorrect, and omitted words manually across several different tools - a slow workflow that consumed valuable time.
Disconnected Tools
Training relied on six separate systems for grading, feedback, and progress tracking -slowing down workflows and creating unnecessary complexity.
To start, I mapped out the needs and workflows of real users. Since there was no usage data from the current tools, I relied on interviews and observations to understand what information mattered most.
Once I had a clear picture, I began iterating different layouts and information architecture based on actual usage - surfacing the most important data upfront and identifying opportunities for automation.
The solution? Automate grading by comparing the agent’s captions to correct transcripts. This automation allowed trainers to immediately view detailed results without the need for manual intervention.
Automating the grading workflow reduces 20 minutes of manual effort per call. Combined across trainers and supervisors, this results in approximately 235 hours saved per week.
A scalable way to match calls to skill level
As we explored the training flow, we realized the real issue wasn’t just outdated assessments - it was the lack of a clear, structured path for agent progression.
Unclear Call Selection
With hundreds of test calls and no way to gauge difficulty, trainers struggled to know which calls to assign - often overwhelming or under-challenging agents.
Through interviews, trainers told us that figuring out which test call to assign was a constant guessing game. With hundreds of calls and no defined way to gauge difficulty, it was easy to overwhelm or under-challenge agents.
To solve this, our UX researcher and I worked with trainers to identify the three key variables that make a call more difficult:
01
Speed.
How quickly the caller speaks.
02
Vocabulary
The complexity of language used.
03
Audio Quality.
The clarity of the call audio.
To address this, we introduced a Clear Call library - a scalable collection of test calls categorized by difficulty. Each call was assigned a level from 1 to 10 based on the different difficulty variables.
Turning progress into a measurable score
While our earlier solutions tackled many training challenges, a key gap still remained: agents often received mismatched training. Without insight into an agent’s skill level, users would struggle to decide which Clear Call to assign.
Lack of Performance Visibility
Trainers couldn’t track detailed progression over time, making it hard to deliver personalized training or identify growth opportunities.
It was during one of our brainstorming sessions - likely fueled by too much coffee - that we hit our next breakthrough. The UX Researcher and I often played competitive games, where player rankings adjust based on performance.
We joked, “What if Clear Calls worked the same way?”
Then it hit us: a skill rating.
To address this gap, we designed the Skill Rating system, a dynamic score (100–1000) that adjusts based on an agent’s performance. Agents are assigned Clear Calls tailored to their skill level, ensuring personalized, evolving training. As an agent performs, their skill rating increases or decreases, providing trainers with real-time insight into their abilities and making it easier to assign the right level of Clear Call.
The Skill Rating system not only enhanced trainer visibility but also gamified the learning process, making progress more engaging and measurable. But we realized we needed to go a step further: apply skill ratings to the entire training flow.
We designed Training Sessions, a flexible system that let trainers assign a series of Clear Calls tailored to specific goals. Sessions could be built in three ways:
- Skill Rating Mode: Agents progress until reaching a target skill level.
- Number of Calls Mode: Agents complete a defined number of test calls.
- Call Hours Mode: Agents train for a set amount of time.
These replaced outdated checkpoints with personalized training paths - supporting both new and experienced agents while minimizing time away from live calls.
Testing
Testing our vision helped us reframe the story
Our first step was presenting the concept to Operations and Training teams, as they were the ones most affected by the manual processes and disjointed tools.
As we walked through the features, the head of training said,
“We can tell you really took the time to listen and find solutions to our problems.”
But the head of operations raised a concern:
“If we implement this, I don’t think most of our team will have a job anymore.”
This was a big moment. We hadn’t anticipated that our push for efficiency might be seen as a threat. We quickly shifted our framing - positioning Cadence not as a replacement for trainers, but as a way to free them up for higher-value work like 1:1 coaching and agent support.
Doubt turned to buy-in after testing
Next, we tested the simplified MVP version of Cadence with trainers and team leads. This version focused on the core functionality of creating and viewing graded Clear Calls and sending out Training Sessions.
Skepticism Around Consolidation
Many trainers were unsure one tool could replace the six they were using.
“I just don’t see how this can do everything we need.”
Excitement After Hands-On Use
Once they explored the prototype, doubt turned into enthusiasm.
“This would save me a lot of time.”
Strong Interest in Personalization
The agent skill ratings made training feel more targeted and scalable.
“This would make it so much easier to pick the right calls.”
One size didn’t fit all, so we introduced Pods
During testing, we uncovered a gap: trainers couldn’t manage smaller groups, like new hires, because the experience only operated at a call center level - making it harder to focus and find the right agents. This introduced Pods:
Pods
Allowing trainers to create custom groups of agents for focused tracking. We designed Pods to be as simple as creating a playlist: trainers could select agents, save the group, and easily monitor their progress.
Results
A clear vision helped guide our MVP build
Instead of focusing solely on the immediate need, we prioritized building out a long-term vision for the training experience. This approach created a clear goal for the team to work toward, enabling us to anticipate future challenges and opportunities.
Reflection
By identifying inefficiencies and bringing everything under one roof, we turned a chaotic process into a cohesive system.
If you’re a developer reading this, you might be thinking, “wow, scope creep” - and you’d be right. Midway through, we had to cut 85% of what we originally designed just to hit our deadline. But that didn’t make the work a waste.
Because we established a clear north star early on, validated our direction with users, and aligned the team around what mattered most, we were able to make confident decisions. This resulted in trust from our users, buy-in from stakeholders, and a shared sense that we were solving the right problems.
If nothing else, this project reinforced just how critical it is to define scope early - not just to keep the project on track, but to help developers sleep at night.
