India should prioritise smaller, specialised and resource-efficient artificial intelligence (AI) systems instead of focusing on developing increasingly large models, according to a new white paper by researchers at the Indian Institute of Technology Madras (IIT Madras).
Titled A Frugal AI Framework for India from CeRAI, the paper by the Centre for Responsible AI (CeRAI) at IIT Madras' Wadhwani School of Data Science and Artificial Intelligence (WSAI) proposes a Responsible Frugal AI Framework aimed at maximising public value while minimising computational resources, energy consumption and financial investment.
Released by IIT Madras Director Prof V Kamakoti at the Industry Conclave on AI Governance 2026, the framework presents an alternative to replicating capital-intensive AI development programmes in countries such as the US and China or relying on foreign AI technologies.
What is Responsible Frugal AI?
The framework combines principles of frugal engineering with Responsible AI, advocating systems that meet clearly defined requirements while using the least possible resources without compromising reliability, robustness or accountability.
Prof Kamakoti said that AI systems did not need to be large or comprehensive to be useful.
“AI does not have to know everything to be useful. It has to know the right things, deeply, for the task at hand,” he said, adding that India needed AI that was affordable, useful and sustainable.
Prof Balaraman Ravindran, Head of CeRAI and WSAI, said India could develop an approach rooted in its strengths in frugal innovation, engineering capabilities and digital public infrastructure.
Smaller models and shared infrastructure
The white paper recommends prioritising right-sized, specialised and multilingual AI systems, particularly for public-interest applications. It also calls for shared AI infrastructure and common technology platforms, or 'public AI rails', to reduce development costs and enable wider access across government departments and public institutions.
It advocates multilingual and vernacular-first AI development, with systems tested under Indian conditions, including variations in internet bandwidth, device capabilities, energy availability and institutional constraints.
The framework also recommends using generative AI selectively, only where content generation is essential, and considering simpler alternatives such as rules-based systems, retrieval-based approaches and conventional machine learning where these can adequately address a problem.
Dr Gokul S Krishnan, Research Director at CeRAI, said that organisations should first establish whether AI was necessary for a particular task before selecting the smallest model or technique capable of delivering the desired outcome.
Seven-point checklist for AI adoption
The paper proposes a seven-point checklist for organisations and public institutions considering AI deployment. It includes establishing whether AI is necessary, defining infrastructure and financial constraints, justifying the choice of model and testing solutions in real-world Indian conditions.
It also calls for tracking resource consumption throughout an AI system's lifecycle, assessing its total cost of ownership and ensuring proper documentation, auditing, human oversight and manual alternatives, particularly for decisions with significant consequences.
The white paper was authored by Prof Balaraman Ravindran, Venkataraghavan KS, Dr Gokul S Krishnan, Prof Balaji Srinivasan and Prof Balkrishna C Rao of IIT Madras.