As enterprises look beyond general-purpose AI, vertical foundational models are emerging as a cost-effective way to deliver specialised intelligence across payments, healthcare, agriculture, finance and other sectors securely at scale.
Is enterprise LLM adoption in India going the SaaS way? Not even a decade ago, SaaS majors realised that India’s fragmented adoptive base for software products required dedicated vertical solutions. The same reality is now dawning on the AI ecosystem.
Vertical models like Razorpay’s Vulcan, Fractal’s healthcare reasoning models — as well as the establishment of its Indian business unit — and BharatGen’s domain-specific models for agriculture and the legal domain all point to a new direction.
Earlier this year, Tech Mahindra unveiled an 8 Bn parameter Hindi-first LLM focused on education use cases, and designed for adaptive tutoring, particularly to address the structural English bias embedded in many global AI systems. This may not necessarily be an enterprise use-case, but it goes to show that vertical models are becoming relevant.
Plenty more are likely to hit the market as software companies in particular look to capitalise on their data advantage.
It’s the clearest sign of a shift among software companies to venture into the AI model game. But it’s one thing to bank on vertical models, and another to sell to enterprises on a long-term basis. Enterprises will not completely abandon general-purpose LLMs and their agentic platforms, so the key will be to make the value amply clear. Once again, we are in the same place SaaS companies were a decade or so ago.
BharatGen CEO Rishi Bal says LLMs like ChatGPT, Claude, Gemini and DeepSeek are getting more and more powerful, but this is due to their versatility. This makes them over-engineered for a specific business problem, and in many cases, getting them to cater to specific needs ends up becoming expensive from a token point of view.
Vertical models are trained on the vocabulary, context, behaviour and logic of one domain, whether it is payments in the case of Razorpay or healthcare for Fractal.
Theoretically, these models should consume fewer tokens for the very particular needs of a merchant or a healthcare provider, or, in the case of BharatGen, agritech and legal services. The widely available popular LLMs can be tailored to these needs, but that means employing a team of LLM specialists in many cases or experimenting, which results in heavy token usage. Both are certainly more expensive options than deploying a vertical LLM.
That’s the pitch at least.
Vertical LLMs also solve the problem of access to AI. Smaller domain-specific models can be run efficiently on modest hardware, allowing startups and smaller enterprises to use AI without fine-tuning a frontier model or maintaining the infrastructure needed to serve one.
Under its Param series, for instance, BharatGen has built models including ParamAgri and ParamLegal, trained on Indian agricultural and legal data.
Bal adds that a strong foundational base model combined with domain-specific training is more useful than either approach alone. The base model provides broad reasoning capabilities, while the additional training brings the vocabulary, context and patterns required for those sectors.
That combination is also relevant in healthcare, and even more critical, one might argue. Suraj Amonkar, chief AI research and platforms officer at listed AI company Fractal, says smaller healthcare models built with its proprietary Vaidya.ai model have shown performance on leading healthcare benchmarks that competes with frontier models.
“We are observing that foundation models and the agentic ecosystem that is catered to a single domain can give results that compete with frontier models,” he tells Inc42.
But it’s not just about the language or the context. In many cases, vertical LLMs are necessitated due to the complexity of signals.
Khilan Haria, chief product officer at payments giant Razorpay, says a single transaction between a merchant and a consumer through an ecommerce platform can contain tens of thousands of structured signals, and decisions often need to be taken within milliseconds. Converting those signals into text and sending them to a general-purpose LLM will certainly add cost and increase latency. It may also not necessarily improve decision-making because each brand-new signal has its own unique structure.
ChatGPT, Claude and Gemini are great models but have not been trained on the proprietary payment data that determines whether a transaction should be approved, routed or flagged. Here, the cost of an error is different, too. The typical human-in-the-loop intervention adds friction to a transaction, especially when it comes to fraud detection. It’s a recipe for customer attrition. This is why engineering teams in fintech companies rely on general-purpose LLMs only for their product development, not for the financial layer, which this product enables.
Haria says a vertical model for fintech and payments is a compelling argument because of three major factors where general LLMs fail: large proprietary datasets, the need for real-time decision-making and the need for diverse use-cases within one domain.
For instance, it is not enough for a payments-specific model to just be great at fraud detection or payments orchestration. To be a compelling long-term purchase for an enterprise client, the model needs to have multiple use-cases such that the entire payments stack is covered.
Indeed, when we assess the other vertical models that are currently out for adoption, these factors are definitely at play. The same conditions exist across healthcare, insurance and lending. In each case, the value will not simply come from a better answer to one question. It is the ability to handle several operationally key tasks in the right order in addition to a domain-aware intelligence layer.
For enterprises, the value will come from the shared intelligence being adapted to their data environment, risk appetite and operating processes. The question is, are vertical LLMs ready to do that?
Amonkar says foundation models represent the knowledge and reasoning layer, while agentic capabilities built on top of them can support multiple enterprise use cases. Citing the example of Fractal, he says that healthcare models developed with Vaidya.ai are being used for report understanding, pre-visit automation, care management and insurance adjudication. These applications use the same healthcare knowledge but perform different tasks.
Let’s say an AI engineer identifies a sector that they know very well and wants to build a vertical model that checks all the above boxes. How do they go about building a custom model?
Here are some of the stages, according to the three vertical model makers we spoke to, starting with the simplest and least expensive steps. Do note that these may not always uniformly apply to every domain.
Once again, we must flag that this is not a one-size-fits-all method. It can vary from company to company and from domain to domain because that’s what true verticalisation entails. The resulting pitch to enterprises will be: one domain-specific model can support several business functions.
The cost of building a vertical model depends on the sector, the data complexity and how custom the model has to be.
Ankush Sabharwal, founder and CEO of conversational AI platform CoRover, estimates that adapting an existing 7-13B model can cost ₹40 Lakh to ₹3 Cr over a period of three to six months, and training it further on a large amount of sector-specific data can cost between ₹8 Cr and ₹15 Cr. As per his estimate, building a large model from scratch can cost anywhere between ₹80 Cr and ₹150 Cr.
The wide range of the estimate is due to domain-specific costs that are hard to ascertain in general. But he adds that a significant portion of the up-front investment would go towards hiring skilled talent, while GPUs may account for 20% of the cost or even higher. Regular updates, testing and compliance can add another ₹5-8 Cr every year, excluding usage-based computing costs.
In Sabharwal’s opinion, IndiaAI Mission’s subsidised GPU pricing for Indian companies and lower engineering costs could make such builds 30-50% cheaper in India than similar efforts in the US. However, GPUs are priced globally, and experienced AI researchers are an expensive hurdle to cross. Data licensing and testing systems are other costs that model builders often overlook.
And after all that painstaking work, the payoff may not ultimately come. Enterprises will evaluate models not by how technically advanced they are or the hard work that has gone on behind the scenes, but by whether the vertical LLM can reduce costs, improve decisions and support enough business functions to justify the investment.
Vertical intelligence could be the next big thing, but getting there may be a bridge too far even for many tech giants.
A stealth reasoning model called Ox Alpha appeared without warning on August 20 on OpenRouter and OpenCode under the label of an anonymous third-party provider.
Offered completely free during its preview window, it features a 1 Mn-token context window, support for text, image, and video inputs, a maximum output of 131,072 tokens, tool calling, and structured output. Its capabilities include coding, sustained agentic work and long-horizon software engineering.
Adoption exploded almost overnight. Within days it processed trillions of tokens, drew hundreds of thousands of users on OpenCode alone, and ranked among the most heavily used models on the platforms that hosted it. Early independent coding benchmarks showed strong results, with some runs placing it ahead of several established frontier systems on real-world software-engineering tasks.
Stripe CEO Patrick Collison called it “very impressive,” while developers raced to test its million-token context and multimodal strengths.
Strangely, no lab has claimed ownership of the model yet. OpenRouter describes it only as a stealth model developed and operated by a provider that prefers to remain anonymous during the preview.
Speculation has focused heavily on Chinese firm Z.ai (Zhipu) as the creator based on tokenizer matches and other fingerprints. Alternative theories point to an unreleased Microsoft MAI model or other Western labs, but none have been confirmed.
The lack of a name has left the industry watching and thinking whether Ox Alpha is a deliberate soft launch, a capacity test, or something else entirely. Its brief open window has already demonstrated that demand for models remains high and that anonymity itself can now serve as a powerful lever in GTM strategy.
Founded in 2024 by Sreenadh Narayanan and Maurice Dhar, Mumbai-based Alchemic is an AI-powered market research company that runs customer conversations on behalf of brands. Traditional market research is slow and expensive, with surveys, panels and focus groups often taking weeks to field and analyse.
Instead of static questionnaires, Alchemic’s AI conducts adaptive, interview-style exchanges at scale, then synthesises the qualitative signal into usable findings.
The system interviews customers, analyses those conversations and returns consumer insights in days rather than weeks, compressing the gap between asking and knowing. The model is pitched at brand, product and growth teams that need directional consumer understanding faster than legacy agencies can deliver it.
The startup operates in the global market research industry, estimated at over $96 Bn, a category under pressure to modernise as AI reshapes how consumer understanding is gathered.
As AI transforms research workflows, Alchemic believes scalable, conversation-driven insights could help shift brands from slow, rigid data collection to systems that continuously surface real consumer understanding in near real time.
What prompts and hacks are CTOs, CEOs and cofounders using these days to streamline their work?
Here’s the prompt used by Proshort cofounder Rahul Aggarwal to get suggestions based on an ongoing deal discussions.
“You are a founder-first deal strategist for B2B startups selling to mid-market and enterprise customers. I’ll share an anonymous deal snapshot with account context, stakeholders, stage, activity, notes, objections, and timeline. Treat me as an experienced revenue operator.
Identify genuinely stalled deals, diagnose the real blocker, and determine how to get them moving. Don’t default to follow-ups. Look for missing stakeholders, unresolved objections, weak business cases, competing priorities, internal politics, procurement issues, or lost urgency.
For each deal:
Be commercially aggressive and intellectually honest. Challenge our assumptions. If the customer doesn’t care, we’re chasing the wrong person, or the deal should be disqualified, say so.
Optimize for movement, not activity. Ground everything in the evidence provided.”
Editor’s Note: Some prompts may need to be adjusted by users for best results or may not work as intended for certain users.
[Edited by: Nikhil Subramaniam]
[Creatives by: Varshita Srivastava]
Source: Inc42
