India’s enterprise AI market is shifting towards specialised vertical foundation models, promising cost efficiency and enhanced domain precision amid rising adoption of industry-tailored systems like Razorpay’s Vulcan and BharatGen’s language models.
India’s enterprise AI market is beginning to split in a familiar way. Just as software companies once realised that broad products had to be adapted for sector-specific buyers, AI builders are now finding that general-purpose large language models are often too expensive, too slow or too blunt for specialised commercial tasks. The emerging answer is vertical foundational models: systems trained for a single industry or workflow, from payments and healthcare to agriculture, law and education.
Razorpay’s Vulcan is a useful example. The company says its payments foundation model has been built on nearly 4 billion payments and trillions of data points, allowing it to analyse thousands of signals per transaction and improve approvals, routing, fraud detection and checkout personalisation. BharatGen is taking a similar approach, offering domain-specific models for Indian languages and for sectors such as healthcare, agriculture, finance and law. Fractal has also developed healthcare reasoning models through its Vaidya.ai initiative, while Tech Mahindra has introduced an 8 billion parameter Hindi-first model aimed at education use cases.
The case for these systems is not simply that they are more specialised. It is that they can be cheaper to run and easier to deploy in settings where milliseconds matter or where proprietary data is too complex to feed into a generic model. In payments, for instance, the decisions are often driven by structured signals that do not translate neatly into text prompts. In healthcare, finance and legal services, the same logic applies: the model has to understand domain language, respect workflow constraints and operate with enough precision to be useful in production.
That does not mean enterprises will abandon frontier models altogether. BharatGen chief executive Rishi Bal argues that broad models such as ChatGPT, Claude, Gemini and DeepSeek remain powerful because they are flexible, but that flexibility can make them inefficient for narrow business problems. Fractal’s Suraj Amonkar says healthcare models built with Vaidya.ai are already being used for report understanding, pre-visit automation, care management and insurance adjudication, showing how one domain knowledge layer can support multiple enterprise tasks.
The economics still matter. According to estimates cited by conversational AI platform CoRover founder Ankush Sabharwal, adapting an existing 7 billion to 13 billion parameter model can cost ₹40 lakh to ₹3 crore, while training on large volumes of sector-specific data can push costs to ₹8 crore to ₹15 crore. Building a large model from scratch may cost far more. Sabharwal says India’s lower engineering costs and subsidised GPU access could make such projects cheaper than in the US, but he warns that talent, data licensing, testing and compliance remain major hurdles. In the end, enterprises are unlikely to buy a model because it is technically impressive; they will buy it only if it cuts costs, improves decisions and supports enough of the workflow to justify the spend.
Disclaimer: This article is intended to inform and educate, not to recommend or endorse any financial product, investment or strategy. Please consider your own financial circumstances and seek professional advice where appropriate before making financial decisions.
