Tata Steel has spent eight years building artificial intelligence (AI) into the fabric of its business. Today, more than 680 AI applications run across the steel major’s operations in India. They monitor equipment for early signs of failure, process invoices, and guide engineers through troubleshooting before a breakdown occurs.
Soummo Bose, the company’s chief AI officer, does not call any of it ‘artificial intelligence’. “We call it augmenting intelligence,” he says, reflecting the company’s belief that AI should enhance human capability rather than replace it.
The distinction matters. That journey took a decade of investment in data infrastructure and specialized talent before a single model went into production.
Most Indian enterprises have not made that investment. Getting AI to work was never the hard part. The harder part is everything that comes after—embedding it into legacy systems, redesigning how people work, and proving it is worth the cost.
Flirty business
There used to be a joke among Indian technology leaders. There were more AI pilots in Indian enterprises than airline pilots in Indian skies.
Rajesh Nambiar, president of NASSCOM, the IT industry’s lobby body, still uses the line, but with a different conclusion attached. “We’re now beginning to see many of those pilots move into production,” he says. “That's the good news.”
But the headline tells only part of the story. Nearly half of Indian enterprises—47%, according to a 2026 EY-CII survey of 200 companies—now have multiple AI use cases running in live workflows.
But running AI in one department is not the same as running it across an organization. Data from NASSCOM and IBM’s Institute for Business Value independently found that only 15% of organizations have reached that harder milestone: scaling AI cross-functionally, with unified governance and measurable outcomes across the business. “Many organizations are flirting with AI,” says Nambiar. “Only 15% are actually scaling it.”
Prasanto Kumar Roy, a Delhi-based technology policy advisor and a former technology editor, has a sharper description for what the rest are left with. “Many businesses are now AI pilot graveyards,” he says. “Proof-of-concepts that may have been successful on their own, but created operational friction when forced to scale enterprise-wide.”
Rishi Aurora, managing partner at IBM Consulting, sees another pattern. Some companies, he says, are not even trying to scale. “I call it a tick-in-the-box exercise. They’re running AI programmes mainly to show their boards they are doing something.”
Where it worked
The companies that have crossed from experimentation into operational AI share one characteristic that has nothing to do with the sophistication of their models. They started earlier, invested longer, and built the foundations that allowed AI to disappear into everyday operations.
At Tata Steel, the journey began eight years ago. Its internal platform, the Tata Steel Digital Assistant or TDA, has an unexpectedly mundane origin: when ChatGPT launched in 2023, the company blocked public generative AI platforms over data security concerns and built its own alternative. Today, TDA runs nearly 200 AI agents used by more than 12,000 officers across India and some UK group companies.
The more consequential work sits inside the plants. Predictive modelling applied to critical equipment has improved availability by between 5% and 20% depending on the asset. Invoice processing has reached approximately 95% automation. The employee who once validated every line item now reviews only what the AI flags as uncertain. “Data and talent have been our two strategic investment vectors for the good part of the last decade,” says Bose. “The results you see today are not something we turned around in a year.”
Apollo Hospitals has 20 certified AI algorithms—12 in live clinical use, having processed 5 million interactions (API calls) across its platforms over two and a half years. Its cardiovascular risk prediction tool, AICVD, has been used in 3.5 million patient assessments over five years. A prospective study found that screened patients showed 4.92 times lower mortality, though Apollo notes the analysis is observational, unadjusted for confounding, and currently under peer review.
What distinguishes Apollo is its operational discipline. Each certified system carries a monthly tracker covering uptime, error rates, and change frequency. Value is assessed per 100,000 interactions. One disease-progression tool is projected to generate approximately $9 million in revenue for every 100,000 patient assessments it processes, based on what comparable healthcare systems in developed markets charge for similar diagnostics.
“The differentiator between organizations that scale and those that stall,” says Dr. Sujoy Kar, Apollo’s chief medical information officer, “is whether AI is treated as a governed clinical capability or a procurement decision.”
The impact of AICVD is most visible in the patients it identifies before symptoms appear. Earlier, cardiovascular risk assessment relied on symptoms, ECG findings, laboratory values and clinical judgement. Patients with no obvious warning signs often left routine health checks reassured, unaware of the risk building beneath the surface. When the system flags a patient as high risk, the consultation becomes more proactive. The physician reviews the AI score alongside the patient’s history and examination, then considers additional investigations, preventive medication, specialist referral and more intensive lifestyle counselling.
Trust did not come overnight. When AICVD was first introduced, many clinicians treated it as a decision-support tool that had to prove itself. Confidence grew gradually as its predictions were repeatedly validated through subsequent clinical evaluations.
“Rather than waiting for disease to become clinically apparent,” says Dr. Kar, “the clinician can now intervene earlier with greater confidence, using AI as a trusted decision-support partner while retaining full responsibility for the final clinical decision.”
At Hindustan Unilever’s Haridwar factory, recognized by the World Economic Forum as a Global Lighthouse site in June 2026, an AI-powered ‘Supply Chain Nerve Centre’ coordinates production in real time. Response times have fallen 72%, inventory days are down 24%, and on-time in-full (OTIF) delivery has reached 99%.
In telecommunications, Airtel’s network-level AI has intercepted more than 27.5 billion spam calls. When fraudsters adapted by routing traffic through international networks, Airtel extended the system to cover overseas calls and added multi-lingual alerts in nine Indian languages.
At DBS Bank India, a generative AI content-capture tool used by 1,500 branch employees has cut data-entry time by around 50%. A separate customer service tool has improved the identification of secondary customer signals by more than 40%. More than two-thirds of the bank’s India employees now use generative AI as part of their daily work.
Tech policy advisor Roy points to Air India’s customer assistant, AI.g, as a case of the transition done well. What began as the airline’s Maharaja chatbot has evolved into an agentic system integrated directly into core operational infrastructure. AI.g now handles nine out of ten customer queries.
The sandbox trap
Getting a pilot to work is one thing. Scaling it across the enterprise is another.
“AI can look sexy when the ‘AI guy’ in a company picks up a problem or two and uses prompts to show how quickly he can get outputs,” says Roy. “But they can’t scale that.”
The reason is rarely the model. The hardest part begins after the pilot works, says Vic Gupta, executive vice president and head of product development at Coforge, which implements AI across large enterprises. “Technology is rarely the constraint. Change management, process redesign, ownership, and governance usually are.”
Four failure modes emerge consistently across the companies that have tried.
The first is data. At Kirloskar, the journey of an AI-powered quality inspection tool from pilot to production became an education in what controlled environments hide. The pilot worked. The defect detection model performed well on curated datasets assembled by the team—spreadsheets, isolated databases, carefully selected examples. Then it met the factory floor.
“We encountered a far wider array of defect combinations in the real-world environment than we had captured in our controlled pilots,” says Ashok Jade, the global chief information officer of Kirloskar Brothers Limited. “Many variables were siloed in disparate systems or were not yet digitized.”
The model had to be rebuilt. Jade calls this the ‘sandbox trap’—the gap between data that is tidy enough to prove a concept and data that reflects industrial reality. The quality inspection tool eventually delivered a 50% reduction in inspection cycle times and a 20% to 25% improvement in employee efficiency, but only after that rebuild.
NASSCOM’s Nambiar identifies data quality as the single most common barrier to scaling AI across Indian enterprises. The Kirloskar experience illustrates why a model is only as reliable as the data it has never seen.
The second is portability. At Tata Steel, eight years of deployment produced one consistent finding: models do not travel. “Despite all the hype around plug-and-play AI, our experience has been very different,” says Bose. A model built for one piece of equipment cannot be replicated on another. Every asset operates under different conditions. Apollo encountered the same problem in clinical settings. Its cardiovascular risk tool performs with greater than 90% accuracy in Apollo’s domestic population, but required full recalibration when deployed abroad, where risk prevalence jumped from 8.07% to 16% in Latin America.
The third is process. “You cannot force-fit the process,” says IBM’s Aurora. “You put AI on top of a current process and expect magic to happen. That is not going to happen."
Saurabh Mittal, managing director and head of transformation and data at DBS Bank India, echoes the same point. The greatest value comes from redesigning workflows around what AI makes possible. At Tata Steel, the lesson was identical—deploy without redesigning and you do not eliminate inefficiency, you automate it.
The fourth is people. Kirloskar began with what Jade calls ‘listening tours’—meeting business leaders not to pitch AI but to understand their operational pain points before proposing solutions. Tata Steel runs a ‘Divisional AI Month’ programme, spending a month with each business unit to demystify the technology and build deployment roadmaps together.
“Deploying AI at scale is, at its heart, a change management challenge,” says Jade, “rather than a purely technical one.”
The guardrails
As AI moves from isolated tools into daily operations, one question emerges: when an AI system makes a mistake, who is responsible?
At most organizations, the answer is unclear. IBM’s Institute for Business Value global study of 2,000 technology executives found that 77% say AI adoption is already outpacing their oversight capabilities.
Apollo Hospitals has spent years building a structured answer. Every AI system must pass through Apollo’s published EASE framework—evaluating ethics, adoption, suitability, and explainability across 18 scored criteria before advancing to clinical use. When an output is flagged as incorrect, it enters a formal corrective action process and is recorded in the model's risk file under ISO 14971, the international standard for medical device risk management. Accountability, by design, never shifts to the machine. “The model is a regulated software-as-a-medical device,” says Dr. Kar, “but it does not absorb clinical responsibility. A human decision and signature close every chain.”
DBS Bank India takes a similar approach through its PURE framework, ensuring all data use is purposeful, unsurprising, respectable, and explainable. Rather than auditing AI after deployment, the bank embeds business, risk, and governance teams into the design phase from the outset. At Tata Steel, every AI system must clear three stages before going live—cold, warm, and hot trials—designed as much to build operator confidence as to test technical accuracy.
The underlying approach is the same across all three. “The best performers bake compliance directly into the product lifecycle,” says Roy. “You cannot do AI now and figure out compliance later.”
The destination
Most Indian enterprises still derive AI value from first-phase efficiency gains—faster processing, fewer manual steps, lower error rates. These are real achievements, but they are not the destination.
“If you are deploying agentic AI effectively,” says Nambiar, “the objective should not simply be cost reduction. It should be about changing your business model, entering new markets, reaching new customer segments. Otherwise, if AI is only being used as automation on steroids, you don’t necessarily need AI.”
Apollo Hospitals applies the same discipline to AI economics as it does to clinical outcomes. A model that is clinically accurate but cannot demonstrate a credible economic case does not earn the infrastructure investment needed to expand. At Kirloskar, every AI initiative is tied directly to a performance indicator owned by the relevant business unit—not a technology metric but a business one.
“CEOs really talk about one thing,” says Aurora. “If my revenue was supposed to grow by X using these technologies, have I achieved that business outcome? Nobody talks to me about the how.”
The companies in this story represent the leading edge of Indian enterprise AI. Even they are only at the beginning. IBM’s India study found that 62% of technology executives expect to be scaling AI broadly across their organizations by 2030. NASSCOM’s Nambiar estimates that Indian enterprises have realised only about 20% of AI’s overall potential.
Closing that gap does not require a better model. It demands investment in data architecture, talent, governance, and the unglamorous work of redesigning processes long before anyone can see the return.
- 15%: Indian organizations that have successfully scaled AI cross-functionally with unified governance and measurable business outcomes.
- 200: AI agents run by Tata Steel’s digital assistant. They are used by more than 12,000 officers across India and the UK.
- 3.5 million: Patient assessments that used Apollo Hospitals’ cardiovascular risk prediction tool, AICVD, over the last five years.