After a week I got my actual 1st project, this time I got lucky with an Insurance giant, the fun part being we got paid by them to save their life. This bot is in the POC level done with the testing and now after 3 months it is finally ready to move in production level.
Here is a prototype of the flow I designed for an inbound insurance FAQ use case.
Here are steps I followed as a beginner:
Post Go-Live Phase:
User Research - Learn about the company, the services they provide, type of customers they interact with. Discuss and understand the problem they are facing and what exactly can be curbed by using a voice bot, this helps to determine the use case of the flow. Listen to the agent call recording to understand the process the flow.
Design the draft flow - Based on the call recordings, and client description design the flow of the conversation. The basic flow generally includes - a happy flow which is directly draws the happiest successful flow, a call-back scenario, a Do not Disturb scenario, and other bifurcated possible use-cases from the main flow.
Add Intent/ Entity - Now in the connecting arrows we add the connectors that is the intent of the bot to move from one state to the next. Keep in mind a bot is not human, it does not have a brain, so at every step it needs to pre-directed as to where to go when. Intents are just directions given to the bot in the form of user response to move to the next path. Entities are inputs given by the user such as - date, phone number, time, etc.
Write Prompts - Fill in each of the blocks with the Bot prompts. This step makes the flow easier to read and understand what is happening in each of the bifurcations. Also add small talks to each bot prompt, these are also fillers to users common recurring action during a conversation such as - asking to repeat, staying silent, or continuously saying Hello! Hello! Have a plan B that is small talk prompt for each state to cover such common scenarios.
Finalize the API - API are interfaces which works in the backend to fetch particular data in micro seconds while on the call, to have a better experience. An application programming interface is a way for two or more computer programs to communicate with each other.
Collaborate with Software engineer - Create the design, generate YAML, finalize the flow with all the conditions, metadata, intent validators and then start internal sub testing, and finally start the UAT.
Go-Live:
In this phase the calls will be triggered as per the number agreed upon by the client in the specific language. There was 1 default language, and the call was also triggered in 3 other local Indian languages for customer efficiency and ease in understanding.
Post Go-Live Phase:
Continues call review - In the internal review sessions of the call it is very impactful to understand how the real customers are actually responding to the bot. At this phase we came up with lots of Out of Scope questions, which was not considered during the designing and testing phase. Those OOS intents are one of the most valuable data to train the bot again and again on the new intents and giving a solution on how to answer those questions.
Measuring metrics - Metrics such CRR, IRR are important to understand where my Bot stands. Did it actually help the client, did it reduce their cost, was the bot conversation efficient enough to get the correct action or answer from the user.
Reflect and Document - Documentation is a complete cycle to the process. With each project documentation it is clear for a third party to understand what is going on in the background. Out of Scope and Out of Domain needs to be clearly defined to avoid mistake. Also the logs should define each state and intent and the reason behind using the same.
Comments