Customer Service Bot
During one of my past roles, I joined a team responsible for developing "proof of concept" AI/ML solutions for potential clients.
One of those proofs of concept was for a major Auto Insurance provider, to revamp their existing customer service chatbot. We used the existing structure, content, and categories as a starting place and improved upon every aspect.
​
I was responsible for authoring the new chatbot's responses and restructuring the category architecture. In order to incrementally improve and validate our changes, I designed and ran crowdsourced validation jobs through a data gathering platform called Crowdflower.
Note: "Crowdflower" has since been renamed, and acquired by a company named Appen.
"Trish", the Front Desk Lady.
In order to preserve the confidentiality of our client, and to make the validation tasks understandable to an audience who would likely know very little about machine learning & chatbots, I created a scenario featuring a "front desk lady" named Trish who worked at a pretend insurance company.
The "crowd" was presented with data & category pairs from the Chatbot's training data (but it was presented as questions asked by customers to Trish, the front desk lady), and were asked to answer whether Trish had given a good answer or not.
Here is a sample of that crowdsourcing job.
Note: To preserve discretion & confidentiality, I am only able to provide the above small, redacted sample.