How CoRover handles the pressures of building a bot for


It was a easy question, however not one they might get flawed. So when CoRover was tasked with creating a chatbot for the India Railway Catering and Tourism Corporation (IRCTC) web site, they knew a lot of their success would rely on how nicely they dealt with the ‘refund status’ question — one of the largest in quantity on what was simply one of the busiest websites in India dealing with round 8,00,000 bookings day by day.

“On AskDisha, people can ask by just typing or by voice in Hindi or English. Plus, there are different kinds of refund statuses on, which we have to anticipate,” Ankush Sabharwal, Founder and CEO of CoRover tells about the chatbot his firm created for the Indian Railways ticketing web site.

Refund queries on IRCTC vary from somebody asking about a failed transaction, the place cash bought deducted however the ticket didn’t come, to a easy cancellation. Another refund question generally seen on IRCTC is one the place prospects wish to know the refund standing of an offline cancellation on the web site. Sabharwal says their purpose was to make sure that the AskDisha chatbot might anticipate all these eventualities.

But it was not a easy downside to resolve. Many instances folks will ask for a refund standing in a convoluted method, which once more they needed to pre-empt. The easy straight method can be the place a person asks ‘What’s my refund standing’ and Disha would then ask for PNR or transaction ID to assist with the question. In a barely twisted take, a person may ask ‘Disha is PNR ka refund status kya hai?’

Ankush Sabharwal, founder of CoRover. (Image by way of CoRover)

“All consumers don’t ask different kinds of questions and 70% of the questions are just repeated. We just have to identify those, and work on that. So instead of spending years training the bot, we need to find the 70 to 80 per cent questions and work on those,” Sabharwal explains.

He admits there could be areas the place a bot will get stumped. “We have learnt that there are 150 ways in which people in India can write the word loan, 80 ways for writing the word refund. Data was also required for accuracy in the chatbots.” But he’s additionally very clear that relating to creating a bot, their purpose is to reply max queries, somewhat than specializing in including too many options.

So when IRCTC’s AskDisha was launched in October 2018 in English, it had simply 288 intents, however now greater than 4,000 intents. The new ones have been developed based mostly on the studying from “daily live analytics dashboard”. The Hindi language function was added in February 2020.

AskDisha has dealt with over 10 billion interactions benefiting greater than 178 million passengers until date. It handles 1,50,000 passenger queries day by day and buyer queries. It has seen a peak of 800,000 concurrent customers and to make sure fail-proof scalability and safety, it depends on Microsoft Azure’s Cloud options.

Building chatbots

CoRover, which began in late 2016, has taken a completely different method to building chatbots and has thus far created many together with AskMaitri for Indraprastha Gas Limited which permits customers to report a gasoline leak. It has additionally constructed the AskPAI chatbot for National Payments Corporation of India (NPCI) to offer digital cost literacy round subjects similar to Fastag, UPI. The firm’s AskSarkar bot was additionally one of the winners of the ‘Aatmanirbhar Bharat’ app problem.

Sabharwal says his firm believes in the Chatbot as a service or CAAS mannequin. So the concept isn’t simply to construct a chatbot, hand it over to the buyer and neglect about it. “We decided we will not do any work where we don’t get access to train the bot. Sometimes clients are like you just create the bot and we will service it. It doesn’t work for us,” he factors out.

But a chatbot pushed by synthetic intelligence and machine studying (AI and ML) fashions is simply nearly as good as the data-sets that are getting used to coach it. “We tell our clients to give whatever they have. We will take information from your FAQ list from the public domain, or if you have data from your existing call centre, then you give it to us, any format. We will train the bot, we will take the onus,” he explains, including that if the bot doesn’t work, it displays poorly on them, and never simply their purchasers.

The firm now has over 300 million customers due to the bots it has created. And it’s going past India with a video-based chatbot known as Yukimi in Japan. The video-bot is a component of a kiosk that helps residents in Kibichuo discover eating places of their selection. The firm has plans to introduce such kiosks in India powered by bots with a human interactive factor, although there’s no precise timeline when these will begin rolling out. Sabharwal says the human contact in bots is required and the firm is exploring AR and VR components in future bots.


Please enter your comment!
Please enter your name here