GenAI helped reduce customer support chat costs by 90%: Meesho CTO Barnwal

We continue to be the most downloaded shopping app for two to three consecutive years, he says

E-commerce firm Meesho handles about 100,000 chats per day. About a year ago, 100 per cent of these were handled by human agents. However, the Bengaluru-based firm has now been able to automate about 95 per cent of its customer support chats using generative artificial intelligence (GenAI), said Sanjeev Barnwal, founder and chief technology officer (CTO) of Meesho. In a video interview with Peerzada Abrar, Barnwal said that this has resulted in a 90 per cent cost reduction for the customer support chat. Edited excerpts:

How has Meesho evolved and what are some of the key learnings for you?

From the time we started in 2015, our mission of democratising internet commerce for everyone in India has stayed consistent. Over time, this has manifested in multiple ways. We have continued to expand and cater to a larger customer base to the extent that we now have more than 155 million annual transacting users.

We continue to be the most downloaded shopping app for two to three consecutive years. The scale is pretty great, but personally, what really gives me satisfaction is the impact we have created by making a vast variety of products available at very affordable prices to millions of consumers in the deeper parts of the country. More than half of the consumers that come to Meesho are buying online for the first time, so we have played a significant role in this transition for a large part of the country.

The same story applies to the supplier side as well. We have enabled many suppliers to leverage online commerce as a distribution channel for the first time. Without Meesho, most of them would never have been able to sell online because e-commerce in India was not traditionally meant for selling unbranded products. The real India buys unbranded products, and a large part of commerce even today happens offline.

I think we have created a lot of impact there. In terms of entrepreneurship, it is important to find a mission that is exciting and bold, pushing us to think big and attract the right talent. It is crucial to have the right talent and build the right culture, enabling everyone at Meesho to have an open playground for innovation. Ultimately, if you are not innovating fast enough, you will not capture the market. We have been able to do this consistently over the last nine years. Those are my initial learnings.

How do you view the momentum in the AI space and innovations by Indian startups compared to the global scene?

I think Generative AI (GenAI) itself was a major disruption for everyone. Even for companies that were actively using AI, Generative AI unlocked a new, simpler way of solving complex problems that would have otherwise taken much more time. In simple terms, it has made applying AI much easier across various domains. Many companies that did not previously consider AI as part of their main strategy are now integrating it because it has become essential. If companies do not leverage AI in the mid to long term, others that do will likely disrupt the entire market through better efficiency, enhanced consumer experiences, and cost-effective solutions.

For many, this shift has been a significant activation event, especially for those who were sceptical about AI. For companies already using AI, Generative AI introduced a new approach to solving problems, offering a unified solution for many common issues instead of addressing each one individually. This innovation provides a powerful tool to tackle multiple problems simultaneously.

Currently, the cost of AI remains high, mainly due to the expense of GPUs and the size of the models, which often have billions of parameters. For example, GPT-4 was launched with around 100 billion parameters, and more recent models, like GPT-4 mini, still have tens of billions of parameters. We also see a spectrum of models, from proprietary ones like OpenAI's to open-source models like Meta's Llama. The latest version, Llama 3.1, performs close to GPT-4 mini in many benchmarks. The availability of powerful open-source models fosters innovation, as more people can fine-tune and improve these models.

Overall, while costs are currently high, we observe significant reductions every six to nine months due to advancements in hardware and software. As prices continue to drop, more AI use cases will become financially viable.

What are the AI technology bets that you are making?

The biggest impact for us has been in customer support chat. We have now automated about 95 per cent of our customer support chats using GenAI, which is a big deal. In a very short period, we have achieved this with an improved customer experience. To give you some perspective, we handle roughly 100,000 chats per day. About a year ago, 100 per cent of these were handled by human agents. Customer support chat inherently comes with complexities, such as fluctuating volumes depending on seasonality and sales, leading to peaks in both orders and support chats. This variability introduces inefficiencies in hiring human agents to handle peak times, rather than being able to scale up and down quickly.

Currently, about 100,000 chats per day are managed, with 95 per cent powered by generative AI, resulting in a 90 per cent cost reduction and a better customer support experience. 

Our journey into GenAI started about a year ago, investing in understanding what problems could be solved with this new technology. We achieved this through a combination of in-house innovations. This involves understanding the context of the chat, identifying what the user is asking, finding the right SOP, retrieving relevant data, and then passing it to a proprietary model like GPT-4 or GPT-3 from OpenAI to generate responses. Additionally, we have developed in-house models to detect when the conversation should be closed. This combination of context detection, SOP identification, data retrieval, and response generation using OpenAI's model has enabled us to revolutionise our customer support chat.

What are the other AI-related applications that you are focusing on?

I think this chatbot is a big deal. Most companies have not been able to achieve this level of automation. Beyond this, we have a unique feature at Meesho due to our user base in Tier-II cities and beyond, where vernacular language is heavily used. Many addresses provided while placing orders are in local languages. Standard translation engines, like Google Translate, struggle to differentiate between proper nouns and words that need to be translated versus transliterated. We have found that LLMs (large language models) perform much better in this aspect. Our generative AI-based system now allows users to provide addresses in their vernacular language, and it translates these to English while preserving the context of the address, such as building names. This is particularly challenging due to the diverse types of addresses in India.

Additionally, we are investing in custom in-house transformer-based models to improve geocoding and other related tasks. Address translation is one key area. Another interesting aspect is handling millions of search queries daily, many of which are in vernacular languages. Even those in English or Hindi-English often have unique spellings and descriptions, such as "shirt pant ka kapra" with "shirt" spelt as "S-H-A-T" and "pant" spelt differently. We use LLMs for spelling correction, translation or transliteration, removing redundant keywords, and rewriting them to more frequently used terms. This helps us understand queries better and surface relevant products to our users, processing millions of queries daily.

These are the top three use cases of LLMs for us, along with many other experimental applications.

What kind of edge are these AI efforts going to give you to compete against players such as Flipkart and Amazon?

I think we prioritise delivering the best experience to our users above all else. Our primary focus is ensuring that our users can find the best products at the most affordable prices. We do not concern ourselves with comparisons to our competitors because each platform has different target audiences and unique challenges, making such comparisons unfair.

That said, we believe we have one of the strongest data science teams in the country. We have leveraged AI, particularly generative AI, in unique ways to address a wide range of problems. It is challenging to comment on specific companies because I do not have insights into their achievements or current stages. However, we are confident in our capabilities and the innovative solutions we have implemented.

What are the long-term AI-related projects for you?

We aim to explore how to apply AI across all aspects of the marketplace. To start, we are focusing on enhancing our search experience and improving personalisation. With 40-50 million active products at any given time, we need to surface the best 20-40 products to our users. This involves real-time personalisation and an implicit understanding of customer preferences.

Beyond that, we are working on helping customers make informed buying decisions by providing essential product information in an easily digestible format, especially for those new to e-commerce. For suppliers, generative AI can enhance cataloguing and business operations. We are looking at ways to improve catalogue quality, help suppliers identify suitable products to sell, and understand user feedback to refine their offerings.

On the supply chain side, AI helps in fraud prevention, better address verification, and overall logistics efficiency. In essence, AI is integrated into all verticals at Meesho. We have a dedicated data science team of about 60-70 members working across every area to drive innovation through AI.