Conversational AI projects such as chatbots and virtual agents to help handle customer interactions, have been widely regarded as a revolutionary way to improve customer service and engagement. However, it is critical to understand how to use the technology strategically in different business environments and avoid common pitfalls in order to succeed consistently.
AI is embedded in at least one business function according to 33% of respondents to the recent McKinsey’s survey. At the same time, merely a quarter of respondents said that AI had benefited the bottom line (defined as a 5% boost to earnings).
There are important lessons learned that help identify some of the main reasons that conversational AI projects underperform. We will explore some of them against the backdrop of our successful project implementation experiences.
The most important caveat: it is unreasonable to expect that AI will be able to do what humans cannot. AI can do things faster, process vast amounts of information at fraction of a second, but it still cannot resolve a poorly formulated issue, or ambiguously expressed customer intent. Customers are always right, of course, but tend to come up with unintelligibly formulated problems from time to time. The good side of AI handling it is that robots remain courteous at all times, even where humans falter, and can recognize real intents even when the caller is unsure what is going on. For example, one Spitch voice bot for a telco use case would correctly determine that the intent was fixing a problem with an internet connection, when a customer complained about difficulties with uploading photos on Instagram.
Perhaps the key misconception is that chatbots can "chat" without significant assistance. There are virtually no plug-and-play options here and even out-of-the-box solutions rely on language models that need to be perfected through assisted machine learning to remain accurate. Even ChatGPT, for example, would not be helpful producing factual responses to customer queries without connecting to a specific in-house knowledge base to inform such responses (unless you want to entertain your customers with fancy novel content). In order for conversational AI robots to be effective, they must be implemented as part of a larger strategy for gathering the right data, fine-tuning to improve the accuracy of recognition, and incorporating artificial intelligence (AI) into an enterprise process and knowledge base to enable correct responses and trigger right actions.
A major challenge is that even when a chatbot is powered by the best available NLU engine and the largest language model, the business process automation results may still not meet all customer needs, or the human agent who picks up the conversation eventually may still be unable to resolve the issue. This is because AI is also about contact orchestration, business process optimization, and agent empowerment. The key to success is addressing real business problems to improve efficiency and reduce costs. This should be part of a multi-pronged strategy focused on empowering both virtual assistants and human agents to assist customers effectively. The accuracy of the NLU engine that stands behind the conversational AI is far less impactful as a success factor compared to the business process and orchestration for the resolution of customer issues and intents identified by the AI.
There are also technological limitations. Even with a "low-risk" approach involving the latest end-to-end universal and domain-specific models to automating frequently asked questions (FAQs), chatbots can still misunderstand customers due to the limitations of NLU. The maximum achievable level of accuracy remains at around 95% and there is very little one can do about it no matter how much effort and money are invested.
The epic race for AI supremacy between the tech giants determined by sheer computing power and language model sizes will rage on, but according to Epoch
estimates, at current rates, “big language models will run out of high-quality text on the internet by 2026 (though other less-tapped formats, like video, will remain abundant for a while)”. This is likely to redouble the importance of high-precision voice recognition, but that 4-5% inaccuracies rate will remain. Cutting edge research in this area is already starting to concentrate on tuning models and making them more compact and focused.
While universal content-generation solutions like ChatGPT may be popular and help build trust to conversational AI, they may not align immediately with the priorities of banks and insurance companies' customer service and HR departments. These and other industries have specific needs for handling customer interactions, employee support, and automated call processing.
To overcome these challenges, companies are turning to one-stop platforms in conversational AI with assisted learning and human feedback. These platforms use incoming user data to fine-tune domain-specific models and provide proven solutions for industry use cases. Spitch is one such vendor, with its omnichannel conversational platform and NLU Suite. Spitch offers a full stack of modern conversational AI products, including intent-level conversation intelligence and analytics. By implementing these solutions, companies can enhance the customer experience and increase efficiency.
Conversational AI projects can be a powerful tool for improving customer service, but they are not a magic solution. To be effective, they must be implemented as part of a larger strategy for incorporating AI into an enterprise. Spitch’s experience helps companies to build and implement such a strategy within the framework of long-term partnerships with their clients. Additionally, companies should consider the limitations of NLU and the need to empower and augment human agents with complementary conversational AI to provide the best customer service possible. Spitch's solutions can help companies overcome these challenges, improve the customer experience and reduce costs.
