Conversational AI is the technology behind chatbots, voice bots, and virtual assistants (VA). Enterprises are looking towards conversational AI technology to enhance customer support, and employee and customer loyalty. The adoption is increasing during pandemic times and is only going to accelerate in the future.
While conversational AI technology can address many business challenges, business leaders often find it challenging to take the right set of decisions and initiatives, given the state the industry is in. In most cases, enterprises have tried to acquire daunting NLP and AI skills in-house and struggled to get a quality solution. In a few cases, enterprises have spent a full year and a few million to get the solution that best fits their business need. In a few more cases enterprises are seeing a lower adoption of the solution and therefore discarding in less than 2 years. The industry is also clouded with numerous vendors across the value chain with vertical to horizontal domain agnostic solution providers. This current state of the conversational AI ecosystem is making business leaders think about implementing a chatbot or virtual assistant to be quite a challenge, demanding a new setup and huge budgets to make it successful. But the fact is Conversational AI initiatives are much easier than perceived when implemented right.
In this article, we can see steps to get started with your conversational AI Journey.
Step 1: Proof of Value
Define the primary use case and establish the return on investment (ROI). Be very specific about the use case. For example, customer support is very generic; a specific use-case could be something like – A website chatbot to address the customer queries on order status. The next step is to calculate the potential ROI. There are multiple ways of calculating the ROI and it varies based on the type of use case. In the case of customer support use-case, the best and easy way to calculate the value is by taking the current query resolution cost and multiply it with the number of queries the chatbot or VA can contain. A good quality virtual assistant can contain about 30% of the total queries. Calculating the value of an employee support VA takes a different route as it factors in qualitative factors such as employee benefits and loyalty in addition to query resolution cost.
Validate your use case by presenting to various stakeholders internally and get their feedback. No need of developing the solution at this stage. You can make use of mock screens to present the concept.
Step 2: Develop the right language model
Once the use-case and the value are identified, the value calculated and validated, the next step is to deep dive into implementing a language model.
A language model is a collection of intents and entities that are fundamental components in the structure of a VA. Identifying the right set of intents and entities is important as it determines the quality of the VA going forward. Creating intents can become a complex task involving AI specialists. In an idealistic scenario, intents are identified by doing text mining by analyzing call transcription of historic data. One needs to use more than one data resources, either internally or from an external source, to identify all the possible intents and entities. One also needs to take into account synonyms, shorthands, and slang (for voice assistant) while designing the intents and entities. A typical customer support VA has in the range of 20 to up to 50 intents or even more.
The above-mentioned approach looks time-taking and daunting, demanding new skill sets and resources. This can be fast-tracked by making use of pre-defined skills provided by players like Smartbots that has Industry-specific and use-case specific skill librarirs. Making use of pre-built and tested skills not only reduces the time to deploy but also ensures the highest quality.
Step 3: Design the conversation flow
After the language model is developed, the next step is to design the possible conversation flows. A conversation flow is the set of possible ways in which users might interact with the VA. Technically, this is nothing but a decision tree. As you might have already sensed, that human conversations, especially in the case of a customer support case, soon tend to become complex, resulting in a huge decision tree that becomes close to impossible to maintain. The solution here is to use advanced dialog management systems, that have the in-built capability to handle possible non-linear flows, that you don’t need to program explicitly. What this means is, while the decision tree you see is small and easy to maintain, the same gets translated to a more complex tree in the backend when users interact. The dialog management should have the capability to dynamically scale as it interacts with more and more users fitting in as per the user’s need. Choose a conversational AI technology that has a robust dialog manager to avoid managing a complex decision tree.
Step 4: Test the language model and flow with automated testing
Thorough testing of the VA is essential before moving into a deployment state. The language model and flow testing are different from that of standard software application testing. One needs to test the VA based on a set of parameters that are unique to a VA. The parameters help in determining the quality of the intents, entities, and if the VA can handle flow deviations. The second challenge is to get the right set of test cases and in the required quantity. A good test case gives out the right indications. A general rule of thumb is to have about 2000 test case conversations that cover various types of conversations like – the happy path conversations, non-linear conversations, conversations that have flow discontinuity, out of domain conversation, and small talk conversations.
Dialogscore is a good tool that comes with pre-built test cases, test-case generation ability, and metrics to evaluate the quality of a VA. The metrics plus recommendations in dialogscore help in improving the quality of a VA. And its ability to quickly generate and run test cases speeds up your VA development time without compromising on VA quality.
Step 5: Integrate with enterprise systems. Ensure human in the loop is included
The next and easy step for enterprises is to integrate with the existing enterprise systems. Most of the enterprise systems these days facilitate third-party integration capabilities, leveraging which simplifies the integration task. If one needs to integrate with a legacy system, or a homegrown system, the integration complexity increases. A way out, in this case, could be to break the project into phases and integrate systems that have in-built integration functionality in the first phase and move to other systems in subsequent phases.
The human in the loop feature is advised to be included from day 1 to ensure fallback handling.
Step 6: Deploy once and use in multiple channels
Build the VA such that it is channel-agnostic. By this, you can leverage the VA across all the channels where your customers interact without the exact development or maintenance effort. Deploy the VA in one channel to start with and slowly scale into all the possible ones.
Step 7: Ensure continuous training pipeline is in place and is easy to use
The above steps help in deploying a high-quality and robust VA. The next step is to ensure the VA scales as per the usage. You need your tools to monitor and maintain the VA. The tool should be intelligent enough to simplify your job of training by providing a simple no-code layout to train the VA. In general, the job of training a VA should not take more than 1 hr per week and should be done by a business analyst with the right training.
The business benefits that Conversational AI brings in when implemented the right way, are significant. Conversational AI is a new technology and has steps that are different from that of standard web and mobile application development or deployment. When the above-mentioned guidelines and best practices are followed Conversational AI becomes easier to implement in your organization.