The rise of vertical AI-native services in India

In October 2025, we published our roadmap on how AI will reinvent IT services. We predicted that AI-native challengers would disrupt this $264 billion industry, arguing that traditional IT incumbents, most of them publicly listed, would struggle to adapt to a fundamentally new technology paradigm.

Nine months later, the market has moved faster than even we anticipated. In May 2026, the NIFTY IT index, the sectoral benchmark tracking India's ten largest IT companies on the National Stock Exchange, had fallen 39% from its December 2024 peak, one of the sharpest corrections in the sector's history. The selloff reflects a broader reckoning: global markets are pricing in the risk that AI will challenge the labor-intensive delivery model that India's IT industry was built on.

The incumbents have responded, but not convincingly. Most have repositioned themselves as AI-first without the operating or financial metrics to prove it. Meanwhile, pressure is coming from an unexpected direction. Frontier labs have concluded that enterprise adoption is the binding constraint on AI's impact and are now building their own deployment businesses, moving directly into territory that once belonged to the incumbents. The implementation gap is no longer waiting to be closed by the IT services giants who own enterprise relationships. The labs are closing it themselves.

The disruption of India's IT services industry is a threat to incumbents as well as a massive greenfield opportunity in a generation. Here we give an update on our roadmap thesis: what incumbents got wrong, what the frontier labs' move into deployment actually signals, and why AI-native companies, particularly those solving deep, vertical workflows, are best positioned to own the next era of enterprise services.

The incumbents responded. But it may be too little too late

The recent selloff has been accompanied by a wave of consolidation. For instance, Capgemini acquired WNS, HCL invested in Sarvam, TCS acquired Coastal Cloud and ListEngage, and Coforge acquired Encora (an enterprise engineering firm.) While incumbents have been quick to rebrand as AI-first, few have shared the operating or financial metrics that would demonstrate AI has structurally changed how they deliver, price, or margin their work. Until those numbers are released, the AI-first narrative remains a story about ambition rather than proof of reinvention. 

Incumbents have seen tepid revenue growth

The performance data is hard to ignore. FY26, which ended in March 2026, was one of the slowest years on record: the top five IT services companies posted an average constant-currency revenue growth of just 1.2%. Revenue per employee has remained flat, continuing a trend that has persisted for a decade.

Beneath those headline numbers lies a more fundamental problem. Most incumbents are still structurally tied to project-based, headcount-linked commercial models. Deal wins are announced as total contract value on multi-year programs that must be re-won, re-scoped, and re-staffed with each cycle. Unlike SaaS, there is no installed base or  subscription layer that renews by default and compounds independent of headcount. Every dollar of next year's revenue must be re-sold.

This model is precisely the inverse of what AI makes possible. The most valuable AI deployments are deeply embedded in the systems and workflows of an organization. They compound over time, improve with use, and don’t need to be re-staffed to grow. However, the commercial model that built India's IT industry assumes the opposite: that value is delivered by people, billed by the hour, and renewed by the relationship. These assumptions are becoming obsolete, and the incumbents most dependent on it are most exposed.

Why frontier labs are now building services companies

The strongest validation of this capability gap came from an unexpected direction. As model capability have progressed faster than enterprise adoption, OpenAI, Anthropic, and Sarvam have reached the same conclusion: the biggest constraint on AI's impact isn't model intelligence, it's embedding that intelligence into enterprise workflows. Each frontier model has moved aggressively to solve this challenge themselves.

Within weeks of each other, the most valuable AI companies in India and beyond announced parallel moves into enterprise deployment. For example, Anthropic launched an AI-native enterprise services firm backed by Blackstone, Hellman & Friedman, and Goldman Sachs, with a model built around embedding applied AI engineers directly inside client organizations to rebuild workflows using Claude. OpenAI followed with the OpenAI Deployment Company, a new vehicle backed by more than $4 billion from 19 investors led by TPG, and moved to acquire Tomoro, an applied AI consultancy, to seed it with roughly 150 deployment engineers and specialists.

Simultaneously, partner ecosystems are extending the reach of frontier labs into the enterprise. OpenAI's Frontier Alliance pairs its forward-deployed engineers with partners including McKinsey, BCG, Accenture, and Capgemini. Anthropic is pursuing the same approach through partnerships with Deloitte, Accenture, and Infosys, with particular emphasis on regulated and complex industries. Sarvam AI partnered with YCP India to accelerate enterprise AI adoption. By pairing Sarvam’s full-stack sovereign AI platform with YCP’s deep cross-sector expertise, the partnership deploys the technology within the context of unique enterprises, operating within their realities, regulatory requirements, and legacy systems within the region.

These AI leaders did not build new deployment entities because the incumbent IT services giants lacked enterprise relationships or delivery scale. They built them because the incumbents lack the AI capability and talent density to close the implementation gap at the speed the labs require. 

For India's IT industry, that is the sharpest indictment yet: the companies best positioned to deliver enterprise AI at scale looked at the incumbent global systems integrator (SI) landscape and concluded the fastest path to deployment runs around it, not through it.

A rare opening for AI-native challengers

As enterprises move from pilots to production, they are making three distinct choices: build in-house, build with a global SI, or buy from an AI-native specialist. The path enterprises end up taking depends on their engineering depth, technology strategy, and the complexity of the use case. Each choice has different implications for where value accrues and where the opportunity for challengers is largest.

Approach #1:- Build in-house 

Digital natives and companies with strong engineering teams are taking the stack into their own hands. Stripe built its own payments foundation model trained on tens of billions of transactions, lifting its card-testing fraud detection rate at large merchants from 59% to 97%, a result that would previously have required a third-party vendor like FICO's Falcon. 

Klarna deployed an in-house AI assistant that now handles a large share of customer service interactions, doing the work of roughly 700 employees. As AI coding tools lower the barrier to deployment, what once required a large offshore delivery team can increasingly be done by a small internal team. This shifts demand away from SIs and AI-native services firms and toward the infrastructure layer underneath: coding assistants, agentic developer tools, model-serving platforms, orchestration frameworks, and the data infrastructure that feeds them.

Approach #2 : Build with a global SI

Many large enterprises want to own their software, but lack the engineering depth to build it themselves. This is the client base incumbent SIs were built for, and they are still winning significant mandates. Accenture nearly doubled its generative AI bookings to $5.9 billion in FY25. In its engagement with Telstra, it took on much of Australia's largest telco's data and AI roadmap, consolidating 18 vendors down to two and rebuilding core processes using agentic AI. This example illustrates the scale and stickiness these relationships can reach.

But this pool of opportunity is shrinking, for three compounding reasons. Simpler use cases are increasingly being built in-house with tools like Claude Code. For more complex ones, AI-native vendors are productizing vertical solutions that are faster to deploy and cheaper to run than a custom SI build, making "buy" the rational default for a widening set of workflows. And the talent equation is inverting: engineers capable of building cutting-edge AI are choosing AI-native companies over incumbent SIs, so the capability gap that once pushed enterprises toward the SIs is now opening up inside them.

Approach #3: Buy from an AI-native specialist 

When a specialist has already solved a vertical workflow end-to-end, buying beats commissioning a slow, headcount-priced custom build. The enterprise gets production-grade capability in weeks, tested across prior deployments, at a lower cost to serve. This option is most compelling where the workflow is complex, regulated, and domain-heavy. For example, we see this in vertical niches such as pharmacovigilance, medical coding and claims, AML and KYC, regulatory writing.

But the lines between software and services are blurring. Enterprises are not choosing between a tool and a vendor, but ultimately how to deliver outcomes. For a widening set of complex, vertical workflows, buying from an AI-native specialist is becoming the rational default, and that is where the most durable value for challengers will be built.

Four business models, all capable of software margins

Four business models are emerging for AI-native vendors, each reflecting a different answer to two essential questions: 

1. Does the customer buy a tool to operate or a finished outcome? 
2.  Is it delivered as software alone or software plus people?

Business models of Vertical AI-Native solutions
	Business model	Definition	Example	Human involvement
1	AI copilots	Software-led, vendor sells a tool. The customer does the work; the product makes them faster.	Cursor is the clearest example: a general-purpose AI code editor where the model autocompletes, edits across files, and acts on instructions while the engineer reviews and accepts each change.	The customer stays in the loop and retains responsibility for the output.
2	Autopilots	Software-led, the vendor sells an outcome.  The software owns the workflow end-to-end and is accountable for the result.	Kintsugi puts sales-tax compliance on autopilot, monitoring where a business owes tax, calculating liability, registering, filing, and remitting, all without a human in the loop.	The customer is not using a tool to help them comply. The software meets the obligation for them.
3	Platform plus delivery	Services-led, the vendor sells a tool.  A platform the customer operates, implemented by the vendor's own team.	Wonderful AI lets customers license its platforms while forward-deployed engineers embed on-site to configure them against the customer's data and workflows across functions such as support, sales, HR etc.	The customer ends up running the platform. The vendor's delivery team is what makes it work.
4	AI-native services	Services-led, the vendor sells an outcome.  The vendor does the work and hands back the result.	Crosby is a vertically integrated AI law firm that reviews and negotiates commercial contracts using AI alongside its lawyers, taking professional responsibility for the output.	The customer hands over the work and receives finished, signed-off contracts. They are buying a result, not a tool.

These models are not mutually exclusive, and the boundaries between them will shift as AI capability improves. Coding assistance is a copilot when the engineer drives every decision, but it becomes an autopilot once an agent can complete the task independently. As software takes on more of the work, companies tend to move across the grid, converting delivery effort into reusable products and gaining margin leverage as they go.

The instinctive read of this framework might be that autopilots are the place to be. Software that owns the outcome with no humans in the delivery loop looks like the purest, most defensible, highest-margin business, and anything with a services layer looks like a concession to labor economics. But that intuition is wrong as margins don’t track the delivery model - they track the complexity of the work.

Palantir is the counterexample that settles the question. It is a services-led, platform-plus-delivery business with forward-deployed engineers embedded in every significant deployment.The company earns gross margins north of 80, because the work is complex, deeply embedded, and hard to replicate. 

The opposite is also true: removing people from the delivery process doesn't guarantee software-like margins. Basic AI customer-service chat is becoming commoditized, driving prices and margins down despite requiring little human involvement.

Five levers to achieve software-like gross margins with aI-native services 

The path to software-like margins for AI-native challengers is possible through  five compounding levers, each of which reduces the human effort required to deliver an outcome, while increasing the value the customer receives.

1. Own the workflow 

Automate one slice of a workflow and you are a feature. Own the end outcome and you are infrastructure. The distinction matters for retention: a customer relying on you for a result is far harder to displace than one using you to speed up a task they still own themselves. 

Graph AI applies this in pharmacovigilance. Rather than accelerating individual steps, it owns the entire regulatory affairs workflow (e.g., case intake and processing, aggregate reporting, signal detection, and risk management) enabling rapid validation and submission of completed safety reports to country-specific regulatory databases . What it delivers is the regulatory outcome the customer is accountable for, not an input toward it.

2. Build the learning loop 

Every engagement should enable the next one to require less human effort. Every exception a person resolves should become a signal the system absorbs. Wonderful's customer-service agents are built this way. Each case that requires human intervention feeds back into the model, so the share agents resolve unaided climbs over time and the human cost per deployment falls with it. The business gets cheaper to run as it grows.

3. Drive down the cost to serve

Model choice, routing, caching, and workflow design should keep the cost of delivering each outcome falling as volume scales. A high-volume agent business routes the straightforward majority of requests to smaller or fine-tuned models, reserves frontier models for the hard tail, and caches repeated work, so cost per interaction keeps declining even as throughput rises. Margin expansion becomes a function of scale rather than a ceiling imposed by it.

4. Reuse across deployments 

A platform serving many customers spreads the fixed costs of integrations, security, compliance, and model operations across the entire base. Each new deployment draws on what was built for the last rather than starting from scratch. Aivar runs this way: its delivery accelerators began as individual client engagements and were distilled into reusable IP, so each new deployment is increasingly an assembly of existing components rather than a custom build. The marginal cost of a new customer falls with every customer added.

5. Deliver with a small, high-skill team 

An AI-native engagement requires far fewer people than a traditional services build, and the people it does require are highly skilled. This is the structural opposite of the large, junior-heavy delivery pyramid that defines incumbent IT services. Skill density, not headcount, is what carries the work. Tessera demonstrates the model: it delivers ERP modernization programs that once required large consulting teams using a small, AI-native team instead, which is precisely what allows a services-led business to run at software-like revenue per head.

How to win as a verticalized player: A playbook for founders in AI-native services 

Foundation models are more accessible than ever, open-source alternatives are improving rapidly, and the cost of building a compelling demo keeps falling. A polished proof-of-concept built on commodity models is no longer a differentiator, it truly is table stakes. 

The verticalized companies that build durable positions do so by making themselves genuinely hard to displace. 

For the founders building, here is what that looks like in practice: 

1. Solve a problem that is hard enough to matter 

The workflows that are easiest to automate are also the easiest to copy. Durable positions are built on problems with high complexity, dense edge cases, and real consequences when they fail, the kind of terrain where a thin wrapper on a foundation model will not hold. 

dCortex operates in exactly this environment, running AI agents that reason, coordinate, and act across airline operations in real time, where decisions are continuous, conditions change by the minute, and outcomes cannot be predefined. Complexity is not a cost of operating in this market; it is the moat.

2. Embed domain knowledge a competitor cannot quickly reconstruct 

The lasting advantage in vertical AI is not model quality, any well-funded competitor can access the same frontier models. It is the proprietary domain knowledge layered on top of them. 

EvenUp's personal-injury platform is trained on hundreds of thousands of injury cases and millions of medical records, a body of structured legal and medical knowledge that a rival cannot assemble overnight. That depth is what allows the platform to draft and value a claim to the standard a personal-injury firm will stake a case on, and what makes it genuinely difficult to replace.

3. Build on top of systems of record, not against them 

The fastest path into an enterprise is through the systems it already runs, not by asking it to replace them. 

Abridge does this inside Epic, the electronic health record that nearly every large US health system operates in: as the first health IT vendor accepted into Epic's Partners and Pals program, its product runs natively inside the standard Epic workflow. A clinician records a visit and the AI-generated note appears directly in the patient chart without the clinician ever leaving the system. Distribution through the system of record is what allows Abridge to spread across an entire health system in a way a rip-and-replace product never could.

4. Own the regulatory and compliance layer

In regulated industries, compliance is not a feature, it is the price of admission. A company that turns it into a structural differentiator becomes extremely difficult to remove. 

Graph AI has done this in pharmacovigilance, one of the most heavily regulated workflows in enterprise: its platform is built around the regulatory frameworks and submission formats that govern global drug safety, enabling filing of completed reports with regulatory authorities like FDA and EMA. Once a vendor owns that submission path and the audit trail behind it, switching is becomes a regulatory risk a drug-safety team will not casually take.

5. Get speedy to compound the lead 

In a market where enterprises are making long-term vendor decisions right now, being in the room early is often decisive. Product velocity and go-to-market speed compound in ways that are difficult to reverse: early deployments generate proprietary data, that data improves the model, a better model wins the next customer, and the cycle repeats. 

The companies that establish themselves as the default solution in a vertical today are building a data and distribution advantage that later entrants will find increasingly expensive to overcome. In vertical AI, speed becomes a structural moat to lay claim on product innovation and market leadership. 

If you’re a founder or team building in the vertical AI-Native services world, we would love to hear from you. Reach out to the team at india_ai@bvp.com.