Opinion | How to make smart bots work very effectively
2 min read . Updated: 09 Jun 2019, 06:58 PM ISTUnfortunately, success with conversation agents for enterprises has been limited
Chatbots, or conversational AI (artificial intelligence), have found quick adoption in enterprises as they help keep customer communication channels open 24/7. Enterprises can also provide customer delight by offering instantaneous responses. Unfortunately, success with conversation agents for enterprises has been limited. The reason: So far a lot of the available systems make building the first chatbot easy but they do not scale-up when the scope of the chatbot or the complexity of tasks is increased.
To be useful and effective, enterprise chatbots need to address two fundamental challenges. First is the robust modelling of domain(s) for which questions are to be answered. Second is the seamless handling of answering questions, taking actions and knowing when to fall back to live agents.
Robust modelling of domains: The case for deploying conversation agents in areas like claims processing or for loan initiation is strong, and chatbots can reimagine the way each of these verticals do business. However, each domain is different and the need is for more realistic conversational AI that is able to handle nuances of the user language and AI that can handle domain and context understanding.
Enterprises should make use of their ready gold mine of data generated by existing customer care support channels where human agents field customer calls. These human-to-human conversation logs (h2h logs) provide unparalleled customer insights that AI tools can instantly analyse and then can be used for modelling the domain.
Many bots fail to scale up to more complex conversations because of a lack of end-to-end fallback plan with continuous learning and improvement.
Once a bot begin to interact with customers, the interactions (h2b logs) can be used to evolve the bot further. AI techniques can be used to automatically discover “failed conversations". Comparing them with h2h logs helps understand how human agents handled a similar topic more successfully in the past.
Using customer insights to drive automation: Conversation provides a natural and intuitive interface for executing the automation of repetitive tasks. Historical data shows that customers usually come to enterprise chatbots to ask three types of questions. First, transactional questions like “What is my account balance?". Second, informational questions like “How do I set up direct bank transfer?", and third, action-oriented questions like “Can you help me transfer $100 to account Y?"
Each needs different information and approaches. Action-oriented questions, which require the chatbot to do something for them, are amenable to automation and help in reducing man hours. In addition, enterprise conversations may start with transactional or informational questions and end up in requesting an action. A chatbot should be able to understand these complexities and respond appropriately.
For this the conversational interfaces need to support a four-stage process: natural language understanding; information solicitation and validation of the elements needed to perform that action; integration with back-end automation, which can be called for executing the action and; error handling and agent hand-off when automation fails. That’s the only way chatbots can effectively be on the job 24/7 and keep the business humming.
Gargi Dasgupta is director of IBM Research India and chief technology officer at IBM India and South Asia.