As discussed in the previous post, many aspects of conversational AI implementation are dependent on the details. The same applies to the tempo of the setup process. For smooth development, it is in both customer and developer interests that the customers know what they want. Concrete communication between businesses help to prevent additional tasks in later stages. Although, demo of the conversational AI can be built in minutes on the company’s website, we tend to utilize 1-2 months of development time plus one extra month of silent live for testing purposes. We have found out that with this amount of time it is possible to adequately communicate and implement customer’s needs to the AI. Chatbot on a company’s website is a brand extension and its image shaper. This should be done right without rushing.
Today, data is the number one asset. Every chatbot is built on it. Knowing the customer is crucial for developing a truly impactful and useful customer support solution. The first stage of the development phase consists of existing data analysis. The analysis of previous chat or email data will reduce the negative side effects of a “cold start“ and speed up the learning process.
In addition to analyzing, we clean the data, consolidate it to 60-100 most popular topics and evaluate it on 2 axes – solvability and occurrence. At AlphaBlues we have created our own Topic Rating Matrix. This simple but powerful method gives a rough estimate on the impact that an AI solution could have on the company and what topics should be prioritized the most.
Solvability estimates the chatbot’s ability to independently solve customer enquiries. Easily solvable topics will be ranked 3. More complex problems that may require API call will be ranked 2 and questions that require human agent input will be ranked 1. Occurrence indicator shows how often topics come up. Most popular topics will be also ranked 3, less frequent questions 2 and rarely occurring topics 1. After that, solvability and occurrence scores will be multiplied.
To bring a concrete example, take a look at the matrix we have created below. This visual representation of our thought process is a good way to find suitable topics for automation.
For example, telecom company’s new customers might want to find good deals for their mobile plans (What is the cheapest mobile plan available for a new user?). This will be solved easily by chatbot as the mobile plan price options are relatively same for most people. Customers might also want to frequently inquire their invoice (How much is my phone bill this month?). Although this can also be completely automated and answered by a chatbot, a custom API call feature is required. And then there are people who have technical issues with their mobile networks (I have been having issues with my mobile signal. What could be the problem?). There can be numerous reasons for a weak mobile signal. In some cases, a common fix will do the trick, but the majority of similar problems require attention of a real human being.
For best results, one should focus on topics that belong to the green area and minimize the number of red topics. Conversational AI projects can be time consuming and resource intensive, so it is vital to immediately get a good overview of the overall concept. This is the reason why Topic Rating Matrix is a useful tool in topic prioritization and discovery.
When the main topics have been mapped out and answers formed, the building process can begin. This stage consists of the evaluation of the answering logic, language understanding and integration to the customer systems through an API. We understand that the content creators inside the company may not be computer scientists, so a simple non-technical bot building tool is also available. When the core structure of a conversational AI is in place, the AI trainers can start their work. The training application can be used by the company’s own people too, if this is their desire. AI training is usually done by analyzing customer enquiries and either validating or correcting AI’s thought patterns. In order for the bot to understand the meaning correctly, every topic should be linked to at least 20 unique phrases (i.e. different ways to ask the same question).
If the chatbot manages to recognize a certain percentage of total intents correctly, you can move to testing phase. It is advised to launch the early final version silently if the company’s website has live chat functionality. Silent launch means that the AI will be implemented to already operating live chat without an official release. The clients of the company will most likely continue to communicate with the virtual assistant just like they did with the human assistant. However, it is important to be transparent and inform the customers who are they talking to. This smoothens the transition from one system to another.
Congratulations! You now have automated a significant part of your company’s customer enquiries with the help of a conversational AI. Unfortunately, the work does not end here. Every continual service or product needs maintenance and support. Testing and training do not end with the launch of the AI. Our world is constantly evolving and people’s demands act in the same way. To keep up with the customers of the company, chatbots require regular care such as bug fixing, AI training and design improvements.
Our previous post revealed that quality data and good training methods are crucial for the success of a conversational AI. Over the years, we have found out some guidelines and best practices to follow when training a conversational AI. More about Conversational AI training methods and the people behind it in the next post!