Sotoon / Case study / Work started in 2020
Contact Center AI
How working with two quality managers changed my starting point: from displaying speech and language outputs to helping a reviewer decide which call needed attention.
My roleCustomer discovery, product scope, reviewer workflow, roadmap coordination, customer testing, and pricing and pilot discussions. The voice and NLP teams owned the models.
This account combines historical product work, an original prototype, and a modern reconstruction using synthetic calls. The last section is my current reflection.
01 / Discovery
Two weeks with the people reviewing calls.
Sotoon already had automatic speech recognition and natural language processing capabilities. I spent two weeks working with two quality managers to understand how those capabilities might fit their work.
The useful question became which call to look at next, and what evidence would help them judge it. Sentiment and keywords could help surface a call, but they did not settle whether an agent had handled it well.
That shifted my focus from a dashboard of model outputs to a prioritized review queue. A reviewer needed the call, its transcript and context, the quality signals, and a way to continue the assessment in one place.
The attention constraint
Which calls reach a reviewer?
Manual review takes time. Sampling calls limits the amount a quality manager has to listen to, but it can also leave relevant calls unseen.
This illustration starts with 10 sampled calls out of 100. The priority counts below are invented to explain the interaction, not historical results or a model evaluation.
100 synthetic calls received
10 manually sampled
90 calls outside the sample
02 / Choosing the scope
Start with quality monitoring.
The broader contact-center vision included automatic answering, outbound automation, classification, monitoring, and quality review. For the initial product, I focused on quality monitoring: an existing job where speech and language signals could help a person find and assess calls.
It meant keeping the reviewer responsible for the final assessment. A frustrated customer could be evidence of a service problem, a difficult situation handled appropriately, or poor agent performance. A sentiment score alone could not make that distinction.
- The system’s part
- Transcribe, surface quality signals, and organize calls for review.
- The reviewer’s part
- Interpret the context, make the QA assessment, and decide on follow-up.
03 / The original prototype
Designed for Persian-speaking QA teams.
The 2020 prototype brought call playback, transcript, call context, and a quality score together. It is an early artifact of the workflow described here.
04 / The review workflow
Try the reviewer’s side of the product.
Select a call, inspect the transcript and signals, then make a review decision. These fictional calls illustrate the workflow; no model is running here.
Quality review workspace
Synthetic reconstruction of the historical workflow. Review changes stay in this browser view.
Review queue
Billing explanation with visible customer frustration.
May 12, 2021 · Agent 42 · Synthetic customer A · Initial model score: 82
Synthetic transcript. Audio is not included in this reconstruction.
Transcript
Customer
I called last week and still do not understand why this charge changed.
Agent
I can compare the current invoice with the previous one.
Customer
I need someone to explain it clearly, not read the same answer again.
Agent
You are right. The difference is connected to the plan change from May 12. I will summarize it and send the billing issue for review.
Analysis signals
Reviewer workflow
Try a review action. Changes stay in this demo and reset when you choose another call.
The system helps with
Transcription · Initial scoring signals · Call prioritization · Topic or pattern extraction
The reviewer decides
Final QA assessment · Score adjustment · Follow-up need · Coaching or operational action
The model changes what reaches the reviewer and what evidence is visible. The reviewer still owns the QA decision.
05 / Taking it to customers
Pricing and pilots were part of the product work.
Alongside the reviewer workflow, I worked on pricing and pilot discussions with sales and customer-facing teams. Interest in the technology was a starting point; we needed to understand whether the operational problem mattered enough for a customer to pay to solve it.
My responsibilities also included coordinating roadmaps across the contact-center, voice, NLP, and engineering teams, defining the MVP, and using customer testing to guide subsequent releases.
06 / Historical results
What the operating results describe.
- Review / reporting throughput
- +35%
- Historical improvement in the operational review and reporting workflow.
- Supporting ASR volume
- 18K+ calls / day
- Daily volume handled through the supporting speech-recognition workflow. This is not a count of CCAI users or reviewed calls.
These are historical team outcomes recorded in my career materials. The synthetic examples on this page illustrate product behavior and do not reproduce a measured test.
07 / Revisiting the work
What I would examine now.
I would look closely at which calls reach the queue, how often reviewers correct the system, and whether suggested actions reduce work or create more review. I would also make each important signal traceable to its supporting evidence.
This is a present-day design reflection, not a claim about features in the original product. The controls below compare the historical workflow with a possible extension, then separate observing, recommending, and acting.
- 01
Audio
- 02
ASR
- 03
Text and language processing
- 04
Quality or topic signals
- 05
Prioritized review queue
- 06
Human QA review
The system interprets and prioritizes evidence before the reviewer acts.
Three levels of responsibility
I would add autonomy only where an action improves the workflow and its consequences can be managed. Classification and summarization do not require an agent.
Observe
Useful AI does not need autonomy.
- Transcribe
- Extract topics
- Identify quality signals
- Organize evidence
Continue reading