Think of enterprise AI right now as a Formula 1 engine bolted to a delivery van. The engine is extraordinary, the chassis is two decades old, the wiring was last touched by a contractor who retired in 2014, and nobody at the depot knows whether the fuel system will hold. Most American companies that bought the engine are sitting in the car park wondering why it won’t move. That gap, between what the model can do and what the business actually runs on, is the prize Indian IT has decided to chase.
The unglamorous work is where the margin lives.
The conventional story about AI’s economic spoils goes to whoever builds the best model. OpenAI, Anthropic, Google, the labs racing toward something like reasoning. Pick a winner, buy the stock, wait. What that story ignores is that almost none of the money sitting in enterprise IT budgets flows to model builders. It flows to whoever can make the model survive contact with a 1990s claims-processing system, a compliance regime written for paper, and a workforce that was trained on the old way of doing things.
The 95% failure rate is not an AI problem
Corporate America is experiencing a quiet disaster: the vast majority of generative AI pilots are failing. Not because the models are bad. They fail because of flawed integration and a learning gap between the tools and the people meant to use them. The pilot works in a sandbox. It dies the moment it has to read a real customer record, talk to a legacy ERP, or pass through an internal compliance review.
The pattern shows up consistently: most executives are experimenting with AI, but a significant portion admit their data and technology are not ready for what they are trying to do. That is a confession, not a forecast. The bottleneck is not the model. It is everything around the model.
This is the deployment layer. It is where the work of actually changing how a company operates has to happen, and it is the layer that has been quietly hidden from the AI conversation for three years. Founders pitched capabilities. Investors funded capabilities. Boards approved capability budgets. Almost nobody funded the slow, awkward task of rewiring the plumbing underneath.
Why Indian IT thinks this is their moment
For thirty years, firms like Tata Consultancy Services, Infosys, Wipro, HCLTech and Tech Mahindra have done the work nobody wanted to do. They wrote the integration code, they maintained the COBOL that was supposed to be decommissioned in 1999, they ran the help desks, they staffed the back offices of global banks, retailers, airlines and hospitals. They know which fields in which database actually mean what they say they mean. They know which manager at which client signs off on which kind of change.
That accumulated context is, in a recent Rest of World analysis, the asset Indian IT is now trying to convert into AI deployment revenue. The pitch is straightforward: a model is only as useful as the system it lives inside, and we already live inside your system.
The IT industry’s real value is the context and understanding of every enterprise’s business and technology landscape, and making the right technology work inside the processes. AI expands that role rather than displaces it.
The technology will keep getting better because billions are pouring in. Enterprise deployment will not. That gap is the opportunity. Clients are running systems so old that the only people who understood them were contractors in their 70s. The challenge is making modern AI work inside those aging environments.
The revenue is real, and the competitor is Accenture, not OpenAI
This is no longer a pitch deck. TCS, in its Q4 FY2026 results, reported annualised AI services revenue of $2.3 billion, around 7.5% of total revenue, up from $1.8 billion the prior quarter.
That is a line item growing at roughly 28% in a single quarter inside a company with nearly 600,000 employees, where the legacy services business grows in single digits if at all.
Infosys is now doing AI work for 90% of its 200 largest clients, with AI services running at 5.5% of total revenue in the last quarter of 2025. Tech Mahindra is selling supply chain optimisation, autonomous workflow, and decision intelligence work into manufacturing, telecom and financial services clients. The addressable market sits at $300 to $400 billion by 2030, larger than the entire Indian IT industry’s current revenue base.
But Indian IT is not really competing with the model builders for that prize. The model builders need them, or someone like them, to get their tools embedded inside Fortune 500 workflows. TCS has partnerships with Google Cloud, Nvidia, OpenAI and Microsoft. Infosys is plugged into Anthropic and OpenAI. The model vendors have made peace with the fact that they cannot do the implementation work themselves.
The real fight is upstream, with the Western consulting establishment. Accenture has made AI the centerpiece of its strategy and has been acquiring aggressively. Deloitte and McKinsey are selling the same story to the same CIOs. IBM disclosed that its generative AI book of business surpassed $12.5 billion in 2025, a number that dwarfs what any single Indian IT firm has yet to report. Everyone is chasing the same budget line at the same client.
And here is the structural weakness: familiarity is not enough. Being close to the systems is not the same as being close to the decisions that matter. No Indian IT firm consistently occupies the upstream advisory role that clients are now explicitly asking for. The AI conversation inside an enterprise does not start in the IT department. It starts in the boardroom, with a slide deck from McKinsey or Bain framing what AI should mean for the company’s strategy. By the time TCS or Infosys gets the call, the problem has already been defined, the commercial terms anchored, and the politically important choices made. The execution work flows down. The strategic margin stays at the top.
The existential bit nobody wants to say out loud
There is a second problem, and it is worse than the consulting gap. Indian IT’s existing business model is built on selling labour arbitrage. Five hundred engineers in Bengaluru to do the work that would cost five times as much in Dallas. The entire offshore industry was constructed on the assumption that high-volume, repetitive knowledge work would always need humans, just cheaper ones.
Agentic AI is designed to eat exactly that work.
The market noticed. India’s benchmark IT stocks index slumped after Anthropic launched agentic tools aimed squarely at automating the kind of repetitive knowledge work that has kept the offshore industry employed for two decades. If Indian IT firms successfully sell their clients on AI agents that replace 500 offshore engineers with one supervisor and a model, they are also pricing their own legacy revenue out of existence.
Infosys, in a presentation to investors, framed the AI transition as different from prior technology shifts the company has navigated. The framing is honest. Previous waves, including the cloud migration, expanded the work that needed humans. This one is designed to compress it. The bet is that the deployment, governance, redesign and continuous tuning work grows faster than the displaced labour shrinks. That bet is unproven, and the firms making it are racing a clock they themselves are winding.
Process debt is the real asset class
The most useful concept is what can be called the accumulated debt inside enterprises. Not just technical debt, the term every engineer uses, but process debt, data debt, and cultural debt. Decades of workarounds, side processes, undocumented exceptions, and people who are the only ones who remember why something is done a particular way.
A model can read a database. It cannot read the institutional memory of a 40-year-old claims adjuster who knows that a particular field in a particular form is always wrong and has to be cross-checked against a second system that was supposed to be retired in 2008. That knowledge is the actual product Indian IT has been selling, quietly, for years. It just was not previously priced as strategy work.
The repositioning attempt is essentially a re-pricing exercise. Take the same context that was billable at $25 an hour as a managed-services engagement and re-price it at consulting rates as AI transformation work. It is the same context. It is a different invoice. Whether enterprise buyers accept the new invoice is the test. Buyers who have spent fifteen years negotiating Indian IT rates down do not easily turn around and pay McKinsey-adjacent fees to the same vendor for what feels, from their seat, like the same relationship. The mechanics of that conversation are brutal: every procurement team has a memory, every master services agreement has a benchmark, and every CFO has a spreadsheet showing what this vendor used to cost. Re-pricing in that environment requires either a genuinely different product or a genuinely different sales motion, and probably both.
What this means for the rest of the AI economy
The capability layer, the part of AI that gets all the press, is commoditising. Multiple labs, similar benchmarks, falling prices per token. Silicon Canals has previously argued that the more interesting question is what AI still can’t do, and the honest answer is that it cannot, by itself, find its way into a legacy SAP instance, a regulated workflow, and a culture that has been rewarded for two decades for doing things a particular way.
There is a parallel architectural debate happening one level deeper, about whether the current generation of models will ever truly understand context. As we’ve reported, the billion-dollar bet on next-generation architectures assumes the current ones cannot get all the way there.
If that view is right, the deployment layer becomes more valuable, because the model will need human-supplied context for longer than the optimists assumed. If it is wrong, and capabilities scale further, the deployment layer compresses faster. Either way, the value migrates from the model itself to the system around it.
Who actually wins
The spoils will split four ways. Accenture and the consulting establishment will take the largest share, because they already own the upstream relationship and are buying the implementation muscle as fast as they can write checks. The model vendors will carve out a meaningful slice through their own professional services arms, because they have to. A new generation of AI-native systems integrators, smaller firms with deeper agentic expertise and no legacy revenue to protect, will take the technically hardest work.
And TCS, Infosys, Wipro, HCLTech and Tech Mahindra will take the rest, which is still a very large number, because they are already inside the systems that need to be rewired and already trusted on the procurement side. The pitch is not glamorous. It does not need to be.
But only two or three of those five will survive the transition with their margins intact. The winners will be the firms that re-price themselves fastest, before their own automation eats the old book. On current evidence, that means TCS and Infosys. TCS has the scale, the partnership roster, and the AI revenue trajectory to fund the reinvention without flinching. Infosys has been the most aggressive at reframing its own model to investors, which is a precondition for reframing it to clients. The other three are running the same playbook more slowly, and slower is fatal in a re-pricing race.
The firms that pull it off will look nothing like the Indian IT of 2020. Smaller headcount. Higher revenue per employee. A seat closer to the boardroom. More like Accenture, less like the back office of Citibank. The ones that don’t will be remembered as the companies that built the engine that automated them.
