As AI becomes the primary interface between institutions and people, one question is becoming impossible to ignore: who does AI truly work for?
At 56th Annual Meeting of the World Economic Forum (WEF), MINDS, their flagship program to recognize organizations that are spearheading meaningful AI-driven transformations with tangible results, accolated Tech Mahindra for advancing linguistic and digital equity through AI. Project Indus, our initiative to build a foundational large language model rooted in Indic languages, was acknowledged for reshaping AI adoption in economies that speak regional dialects, which global systems often cannot address accurately. This highlights an important shift in how AI adoption is evaluated globally, not just for its technical sophistication, but also by its ability to serve diverse populations responsibly.
India’s Own Foundational Model
When we began this journey, the accessibility gap was evident. Most large language models were trained predominantly in English and a handful of high-resource languages. Yet in India, only a fraction of the population interacts comfortably in English. Hundreds of millions communicate in Hindi and other regional languages, many of which have limited digital representation.
Project Indus was conceived to address this imbalance. It aims to preserve and digitize underrepresented languages while creating scalable, industry-agnostic language models that can power real-world applications across banking, healthcare, education, and citizen services. This is a civilizational initiative that has developed India's own foundational large language model (LLM) focused on Indic languages. One of Project Indus's key objectives is to preserve Indic languages and dialects that lack significant digital records. It also supports the creation of localized and verticalized, industry-agnostic Indic LLMs, enabling us to forge partnerships with telecoms, hyperscalers, and OEMs to build a platform offering E2E LLM solutions.
Bringing Cultural Equity with Regional LLMs
Today, many digital public services across healthcare, banking, and government are delivered through AI-powered chatbots and IVRs. However, most of these systems rely on language models trained primarily in English and a small set of Western languages. For example, India has 27 official languages and more than 1,600 dialects, with only 10-20% of the population speaking English. Language models that are not trained in these regional dialects and cultural nuances fail to recognize intent and cultural context and translate inaccurately.
With our sovereign AI model, we are trying to bridge this communication gap by enabling users to engage with conversational platforms trained in regional dialects and cultural sensitivities. Capable of supporting both speech and text, this model can adapt to any conversational interface. A defining aspect of it has been its training on datasets obtained from open national databases, literature, archives, and user contributions. Such a unique LLM allows enterprises to retain complete control over data and model behavior, ensuring privacy and ethical AI governance.
Built from the Ground Up and Enhancing Accuracy for Regional Context
In my view, digital inclusivity cannot be achieved by retrofitting global models with translation layers alone. Linguistic diversity must be treated as a foundational design principle.
Most popular platforms are trained predominantly in English and other major Western languages. These interfaces perform extremely well in English but lose contextual and translational accuracy when used with diverse languages and dialects. However, our LLM has been trained with linguistic diversity at its core and further refined by native speakers, enabling it to achieve 92% accuracy in Hindi. In comparison, the global benchmark was only 70%.
Our Indic LLM today handles 3.8 million queries per month for banking, healthcare, and government organizations. It is a classic case study that shows how AI can be trained and scaled, with people at the heart of its framework. Such an innovative approach has made AI more adaptive, accessible, and preferable, bringing it closer to people.
Conclusion
Digital inclusion will define the next chapter of AI adoption.
As AI systems become embedded in public life, we must ask ourselves whether they reflect the linguistic and cultural diversity of the people they serve. Building sovereign, culturally grounded foundational models is not simply a technological ambition—it is a societal responsibility.
At Tech Mahindra, we see ethical and inclusive AI not as a positioning statement, but as a design imperative. The future of AI will not be determined solely by how advanced it is, but by how equitably it is built and deployed.
And that, I believe, is the true measure of progress.
