Chatbots hallucinate and get confused, says Jio Haptik which is ‘training’ them

• ChatGPT backed superbots are not the holy grail that can solve all problems of consumer companies. It took Jio Haptik six months to learn that these bots need to be schooled before they can be unleashed in the real world.
• LLM backed bots are good for simple queries but are no good when they are expected to multitask.
• For enterprises that want to improve customer satisfaction using bots, Haptik’s Contakt has solved many of the challenges around data privacy and hallucination.

When OpenAI launched ChatGPT a year ago, it was seen as a superbot that could slay any problem in a jiffy. Companies across the world went into a huddle to figure how they could unleash the superbot to address concerns of their respective customers. While companies were looking at ChatGPT’s use cases, many were quick to pronounce the death of CRM companies. Jio Haptik – a leader in conversational AI – also went into a huddle to figure out its own future after the launch of OpenAi’s ChatGPT.

"Surprisingly, Haptik’s team realised that GPT is not a one-size-fits-all solution for all problems." Explains Swapan Rajdev, Haptik’s CTO and co-founder, “When GPT launched, it was perceived as this holy grail that could solve everything. As we started digging deeper, we realised it struggles with multitasking. If you give it a small task and give clear instructions, it will work very well. As more things get added it will get confused."

Haptik decided to figure out a way for enterprises to ensure good customer experience using this new disruptive technology. Actual production grade scalable chatbots are not simple. GPT has ensured that bots can be conversational, but a lot more needs to be done to train them to solve problems in real time. Training a bot that can represent an enterprise means the bot has to undertake several tasks at the same time. One prompt, therefore, cannot answer different queries. Haptik has created an architecture using LLM to address customer engagements not just through the bot but also for contact centre agents who often struggle to understand a customer’s history and context in real time.

What consumer companies don’t realise is that hallucination is a bigger problem than many envision. An LLM powered bot can hallucinate when it is asked a question it has not been trained to answer. Haptik decided to work on solving many such problems that would spring up while dealing with consumers. Privacy is one thing that rarely comes up when companies are looking to roll out LLM powered bots on their websites. If it is not trained properly, a superbot can possibly share existing data of customers too. Jio Haptik has spent the last six months finding ways to train this superbot so that guardrails are built around such LLM powered superbots. It has come up with a product (Contakt) that can help companies engage smartly with their consumers, but without the pain points.

Artificial intelligence has been around for a while now and has been used by companies to address consumer queries. But in the past, CRM companies had to envisage and feed in potential questions and answers into the system. But it is hard to imagine every question that can be asked by individuals. Haptik’s Contakt can allow a sweep of the knowledge management system of companies to come up with a bot that is well trained to represent the company.

Jio Haptik’s CEO & co-founder, Aakrit Vaish, says: “AI has been key to customer experience but limitations of traditional, rule-based chatbots have often left customers frustrated. The launch of ChatGPT was truly a pivotal moment in our history as an industry, and now the technology finally exists to make world class CX a possibility. With Contakt, every brand can now have their own custom LLM assistant that delivers end-to-end customer experiences, starting at the pre-sales stage, driving transactions, and finally in post-sales engagement to unlock customer loyalty.”

Powered by LLMs at every touchpoint, Contakt is designed to enhance the user experience throughout the customer lifecycle, enabling every brand to have a ChatGPT-like solution of their own. Contakt is more than a standalone solution that aims to streamline customer experience. It empowers customers with self-serve options and automation of routine tasks, and improves agent productivity with co-pilot offerings. Further, the suite’s analytics engine helps track bot and agent performance, and optimize for better customer outcomes with AI-powered recommendations.

Contakt is well suited for pre-sales conversational buying guidance and for customer self-service. Its other interesting feature includes Agent Co-pilot for superior agent productivity, and Intelligent analytics to track and optimize metrics that truly matter. The platform’s self-serve support resolves a vast range of customer queries on its own and actively learns from customer conversations to provide natural and adaptive responses each time, enhancing the overall customer experience.Agent Co-pilot equips agents for better productivity and conversational assistance. It helps reduce query resolution time, query handling time, first-response time among other metrics for improved customer satisfaction.The analytics feature on Contakt allows businesses to create custom reports and dashboards, and get a comprehensive view of customer outcomes, including customer satisfaction score, query completion rate, and others.

Commenting on the integrating GPT like capabilities into its sales function, Anand Thakur, Chief Product and Technology Officer at Reliance Retail, says: “The integration of ChatGPT-like capabilities into our web sales assistant will not only streamline the buying process but also elevate overall customer experience - turning window shoppers into buyers! We believe that this will play a pivotal role in shaping the future of retail tech, providing a unique and tailored experience for every Tira Beauty customer." The company is all set to launch its Generative AI Beauty Advisor chatbot on the Tira Beauty website with Jio Haptik.

SEE ALSO: OpenAI exec says this is the most overhyped — and underhyped — thing about AI right now