A founder I spoke with last month had just launched a customer-facing AI agent. His problem was not the model. The model was fine. The problem was that the agent behaved unpredictably in a live customer environment, where even an occasional mistake carried real commercial consequences. His question was not “which model should we use.” It was, “why is making AI work inside a real business so much harder than building it?”
That question is going to define the next phase of enterprise AI in India. By the standard metrics, India is running ahead of the world on adoption. According to Deloitte’s State of AI in the Enterprise 2026 report, 40% of Indian organisations report significant or full AI adoption, compared with 28% globally.
Yet widespread adoption has not automatically translated into widespread business value. Research from MIT Sloan Management Review and Boston Consulting Group suggests that fewer than 5% of companies have been able to create transformative financial value from AI at scale.
For the past two years, the conversation around enterprise AI has centred on capability. Which model is better, which use cases to prioritise and how quickly to adopt generative AI. Those were the right questions when AI was still experimental. The situation now is different. AI systems have moved into production, and the challenge is no longer whether they are smart. The challenge is whether they can operate reliably inside real organisations, across legacy systems, business workflows, compliance requirements and customer interactions.
EY and CII’s AIdea of India: Outlook 2026 reports that 47% of Indian enterprises now have multiple generative AI use cases live in production. Beneath these numbers sits a paradox that ought to give leaders pause. The same EY-CII study finds that 95% of Indian organisations allocate less than 20% of their IT budgets to AI. Only 4% cross that threshold. India is deploying AI at scale on a capital diet no other major economy is attempting.
Traditional software engineering assumes predictability. The same input produces the same output. AI systems do not work that way. Their behaviour is probabilistic and context-sensitive. A prompt change, a model update, a shift in enterprise data, any of these can alter outputs in ways traditional software never would.
In a pilot, an incorrect output is an inconvenience. In production, it becomes a customer complaint, a compliance risk, a governance concern, or a financial loss. This is what most Indian enterprises are discovering right now. Building the model is not the problem. Making the model work reliably inside a real business is.
That distinction matters because it points to a skill set India’s engineering education has not yet built at scale. The engineers who solve this problem are not the ones who train foundation models. They are the ones who understand how a model behaves under real customer load, how to design guardrails that catch failures before they reach a user, how to build evaluation systems that measure whether AI outputs are actually correct, and how to integrate probabilistic systems into deterministic enterprise workflows. Add to that a working knowledge of compliance, security, and the specific business process the AI is meant to serve.
This is a real discipline. And it is what companies are starting to hire for.
Some of the most closely watched hiring signals in AI right now come from OpenAI, Anthropic, Palantir, Databricks and Microsoft. These companies have been quietly building teams of what they call Forward Deployed Engineers, or FDEs. The role sits at the intersection of software engineering, AI, product, and customer implementation. FDEs do not build foundation models. They make foundation models work inside a specific customer’s environment, against a specific customer’s data, under a specific customer’s constraints.
BCG’s Build for the Future 2025 report, based on a survey of more than 1,250 companies worldwide, found that only 5% are generating substantial value from AI at scale, while 60% are seeing little material return despite significant investment. The differentiator isn’t simply access to better models, it’s the ability to integrate AI into enterprise workflows, establish governance, evaluate performance, and continuously improve deployments. These are precisely the engineering challenges that Forward Deployed Engineers are designed to solve.
For India, this is a rare kind of opening. In conversations with more than 1,200 companies across the enterprise landscape, a consistent pattern comes through: firms are willing to pay a premium for engineers who can take a general-purpose AI capability and make it work reliably inside their business. The demand is not concentrated in a handful of tech firms. It shows up across banking, insurance, retail, healthcare and manufacturing.
There is a version of this story that reads like the last chapter of India’s IT services boom, but with new tools. The world builds the frontier. India runs the deployment. If that were the ceiling of the opportunity, it would still be commercially meaningful. But it would also be a diminished version of what is actually available.
The FDE role is not a services role. It is a product engineering role that happens to live close to the customer. The engineers who do it well shape what the AI product becomes. They influence the roadmap. They identify failure modes that model teams never see. They build the reusable infrastructure that determines whether AI deployment scales or stays bespoke. Companies that build strong FDE benches end up shaping their AI products, not just delivering them.
Whether India ends up on the product side of that line or the services side of it depends on a specific choice: whether the country’s engineering education treats production AI as a serious discipline, or as a downstream skill that engineers will pick up on the job.
For CIOs, the implication is clear. Hiring more engineers focused primarily on model development will not, by itself, close the AI execution gap. The constraint lies elsewhere: in the engineering required to integrate AI into production systems reliably, securely, and at scale.
For engineers, the most valuable AI skills over the coming years will extend well beyond model training. They will include designing evaluation frameworks, building retrieval systems and guardrails, integrating human oversight, and monitoring the behaviour of probabilistic systems in production.
For educators, the message is equally clear. Curricula built primarily around model development and generic MLOps no longer reflect where AI engineering is headed. As AI moves from experimentation to enterprise deployment, engineering education must evolve to prepare graduates for the realities of production.
India’s lead in AI adoption is real. Whether that lead translates into lasting competitive advantage will depend less on how quickly we adopt AI than on how reliably we deploy it at scale.
(By Amar Srivastava, CEO-Online and Group CPO, Scaler)
Published - July 21, 2026 08:00 am IST