‘The Bharat is not in this room…it is the vegetable seller who doesn’t know GB’: Inside Indic language scaling at ET AI Conclave & Awards 2025

“The Bharat is not in this room, it is that vegetable seller who doesn’t know GB, who doesn’t even know AI. But it somehow works,” said Gaurav Aggarwal, Chief AI Scientist, Reliance Jio, drawing the attention of the gathering at the ET AI Conclave & Awards to an axiomatic truth, one which is often ignored across boardroom presentations and AI pitches.

In 2022, when OpenAI introduced ChatGPT, the generative artificial intelligence (AI) model, it changed how the world searched. What it also spurred was an “AI arms race”, with the United States and China emerging as the two “AI superpowers”, and the latter launching DeepSeek-R1, an open-source model by a Chinese startup, in January 2025. China’s Mandarin LLMs have already set a linguistic benchmark. Given this context, where does India stand, especially post the 2024 launch of the IndiaAI Mission and its vow to rank among the top three AI powers focusing on “AI for humanity”?

At the ET AI Conclave & Awards 2025 in Bengaluru on February 26, 2026, this question was at the heart of The Multilingual Bridge: Scaling Indic AI panel featuing Gaurav Aggarwal of Reliance Jio, where he was joined by Baskar Subramanian, Co-founder, MD and CEO, Amagi and Safiya Husain, Co-founder and Chief Impact Officer, Karya. The conversation was moderated by Swathi Moorhty.


Together, these panellists spoke about leveraging proprietary data and infrastructure to propel India beyond its tag as the “AI use case capital of the world” towards being a land of empathetic, context-aware models tailored for Indic users.

Why is Indic language AI important for Indian AI?
To answer the question, why AI in Indic languages matters for India’s AI acceleration, one has to look at these numbers: Fewer than 7% of the nations’ 1.4 billion people are fluent in English, which leaves them out of the AI space. Hence, for India, democratisation of AI is not a mere catchphrase. True scale demands models trained on Indic languages, the 22 scheduled languages of the Indian Constitution, such as Hindi, Bengali, Tamil, Telugu, Assamese, and Sanskrit, that resonate with the nation’s cultural nuances, from regional idioms to multimodal interactions via voice and image.

This vision is central to startups and enterprises in the pursuit of amassing high-quality Indic datasets, addressing the “Indic data problem”, which refers to the scarcity of digital content in Indic languages, something that acts as a deterrent in the training of robust LLMs.

‘Another Jio-moment’
Aggarwal redefined scale for Jio’s millions-strong base: “Scale is not just a number... Scale is diversity for us... Dialect is just the next layer, but all of us are very individualistic... when we speak, we have our own language.” He evoked Jio’s data revolution for the masses: “I think you cannot achieve scale in India unless [the] cost is not noticeable... We want another Jio-moment in AI so that... they go away. And we forget that it even exists.”

Jio’s 5G and vernacular apps power voice-first models, mirroring UPI’s seamlessness. Aggarwal stressed evaluations: “We think voice AI for Indian languages is done. I think it is not even close... human perception of accuracy. It is the make or break.”

‘Data is the weakest link’
Subramanian called data the “weakest link,” escalating to video: “Data [points] to curated, good data... The moment you add video... cultural nuances becomes extremely more diverse... our homes, our buildings... holistic 360-degree human connectivities.” Amagi’s cloud ingests regional broadcasts for tonal mastery in Tamil or Hindi.

He pitched “knowledge royalty protocol” with provenance: “Can we build in provenance... this percentage of my probability that I took from here... create an economic model where every content knowledge source... can actually be provided a monetary model.” Globally: “India... has a responsibility... to what I call the global AI... I call it the Nalanda project”, exporting Vasudhaiva Kutumbakam for unbiased models.

‘Are we scaling right?’
Husain qualified scale: “It’s not just about scaling, but scaling well... Is it the right data? Is it from the right users? Across every segment?... What is the sampling strategy?” Karya mobilises rural workers for ethical corpora, debiasing Western models via four eval layers: public, expert, community, workforce.

She rejected binaries: “We pick people as either builder or beneficiary of AI. But this is not an or situation... a virtuous cycle where... they have to be included at the data level... model level... evaluation level... The more people... the better AI will respond to all of our nuances.” Priority: “Investment at the bottom of the pyramid... that’s how everything else... will scale and fundamentally scale well.”

Aggarwal warned against complacency across more than 19,000 dialects; Subramanian eyed provenance economics; Husain stressed on inclusive cycles.

All agreed that native models must solve local queries, from Telugu voice commerce to Assamese health queries, and so on. As Aggarwal closed: “AI is like math... not going to get commoditised... The desire to make machines smarter... is going to stay.”

The ET AI Conclave & Awards 2025 has L&T Finance as the NBFC partner, Snowflake as the AI Data Cloud Centre, EY as the Evaluation Partner, and T-Hub as the Ecosystem Partner, and is driven by BYD, with Indri and celebration partner and Zoho as the Technology Partner. Vahdam India, Vaaree, Natch, and Andamen are the event’s gifting partners.