It’s been three weeks since we launched our bot, and we’ve learned a lot! And by “we”, I mean us humans AND the bot! Recent enhancements include:
- Improved instructions for creating a case
- Revamped our license conversations flows to improve the success rate
- Made matching against login and access issues less “greedy” to improve accuracy
- Improved logic for login and access issues
- Improved training phrases for “request an agent” and “download a product”
- Added a new download conversation flow
And we have more improvements in progress based on the conversations the bot has had so far. When we set out, we trained the bot against an analysis of 2 years of case and chat transcript data, but we’ve been pleasantly surprised by the variety of questions we are getting. Rest assured that we are working diligently to improve the bot’s ability to recognize customer questions and provide the correct answer.
Our system uses a sophisticated machine learning algorithm to analyze conversations based on match rates, completions, abandons, and escalations, and then recommends changes to our conversation flows and intents. These recommendations are reviewed daily by our team at Qlik. We then make updates to content and flow designs, or create all new flows to address new types of questions.
In our first week, we saw a steady increase in the % of questions the bot recognized as we tweaked the original design based on the initial training. In the last two weeks, though, we’ve seen the start of a pattern where the match rate is high on Monday, then declines through the week as volume increases. Then we roll out changes on Friday and the match rate jumps back up! We expect this to continue for a few more weeks as we collect the initial breadth of questions customers expect the bot to answer.
Thank you to everyone who has chatted with the bot, and stay tuned for more updates!
Qlik Digital Support
