Success Stories

Customer Churn

Comcast | Senior Language Engineer‬

Jan. 2021

Problem:

While working on an annotation and ML project using chat transcripts from hundreds of customer service conversations, I noticed that a very common ending to the troubleshooting sessions was that customers quit the chat.

Action:

I decided to conduct some user research on the data, by reading through every chat, employing conversation analysis techniques to determine which speech acts were present and which related patterns most often led to customer churn.

Result:

My results showed that the most common speech act affecting customer churn was miscommunication where agents did not address customers' questions or where agents made customers repeat actions too many times. The length of the chat also had a strong impact on churn; the place where agents spent the most time and most conversational turns was the information gathering stage. This averaged about 10 mins, mostly wait-time, and 12 agent turns. So, I wrote new dialogue to gather the most pertinent information in 1 agent turn, which would take 1-2 mins for customers to complete. I presented the results of the research and the new dialogue was implemented the following sprint.

Where do the customers need to go first?

Aspen Technologies Group LLC | Conversation Designer

Dec. 2022

Problem:

I was asked to review and make suggestions on the flow diagrams for an IVR system that an insurance client's customers would use to request emergency roadside assistance. I noticed that the order of the prompts could be improved by using principles of information architecture.

Action:

I wrote up a draft of the original prompts in order and edited it to fit best practices. The original order was like this: “Thank you for calling Insurance Company. For customer service, press 1; to file a claim, press 2; for roadside assistance, press 3... [a few more options]… Para español, oprima 9…” I changed the information architecture from "most common first" to "high-priority first." This way the language selection options for English and Spanish would precede the list of all available queues in English and the prompt for emergency roadside assistance would come before everything else. Roadside assistance moved from the third interaction to the second for all customers; and was reached with much less wait time for Spanish speaking customers. 

The revised order was like this: “Thank you for calling Insurance Company. For English, press 1; Para español, oprima 2…” Then both English and Spanish queues would have the same order of prompts: “For roadside assistance, press 1; to file a claim, press 2; for customer service, press 3…" Then I edited existing dialogue, wrote new dialogue where necessary, and updated documentation to support my design choices before submitting my work to the project manager.

Result:

The project manager review my changes and appreciated how I saved customers time in getting to their more urgent destinations and necessary queues in their native languages. She submitted my suggestions to the client so I could present them. After the presentation, changes based on my work were added to the upcoming sprint.

Jumbled IVR Prompts

Aspen Technologies Group LLC | Conversation Designer

Feb. 2023

Problem:

My team had created dozens of callflows for our client, using Amazon Lex and Amazon Connect. Our spring demo was fast approaching and we needed to test all of the queues that our client's customers would be using. I was in charge of testing the US Spanish and CA French queues since I had written the dialogue for them as the only polyglot on my team. During the first round of calls, I noticed that the prompts' languages were not mapped to the correct queues and the prompts were out of order.

Action:

I escalated this issue to the lead builder and project manager with specific notes on the proper order for the prompts and ideas about why they might be jumbled. When the lead builder reached out for more information, we did a conference call for another test and I explained which prompts were wrong in real time.

Result:

Next, we investigated the issue in the AWS console and determined that the main queue's initial language selection was not permeating through to the subsequent queues. So, the language selection was changing at each queue, messing up the order of the prompts. The lead builder made updates to the workflows which preserved the language selection value and deployed an update to our dev environment. Finally, I redid my test calls to see if our update worked. My results showed that the languages and prompt order were now correct and permeating through each queue. I updated my notes, reported this info to the lead builder and project manager, and moved my work ticket to the "ready for demo" column of our Jira board.