The Real Truth Behind Indian IT’s ‘Thousands of AI Projects’

During the Q3 earnings season, Indian IT boasted building thousands of AI projects, running hundreds of agents, and shifting to an “AI-first” delivery model. The numbers sound transformational, but the reality is far more layered, uneven, and revealing about whether the industry is truly changing or simply doing the same work faster.

The problem begins with using phrases like “AI-infused”. 

Sanchit Vir Gogia of Greyhound Research describes the current state as hiding behind one comforting phrase. “When providers use the phrase ‘AI-infused,’ they’re often covering a wide spectrum of activity under one friendly label, and that label hides the real complexity,” he tells AIM.

At the base of that spectrum, nothing looks new from the outside. Developers use Copilot to write code faster. Test engineers generate scripts with GenAI. Support teams summarise incidents through chat tools.

“This layer is invisible to the client. The scope of work hasn’t changed, but the delivery model has,” Gogia says. “It’s more efficient, and it protects the provider’s margins.”

Gaurav Vasu of the cognitive intelligence platform UnearthInsight hints at the operations where Indian IT uses to ‘infuse’ AI. These are: application modernisation with AI-assisted coding, automated testing and regression checks, data cleanup and schema mapping, incident prediction, and log analysis in infra ops. 

“These are traditional IT services where AI changes how work is done, who does it, and how fast outcomes are delivered,” Vasu notes. “Less manual execution, more supervision, validation, orchestration, and exception handling.”

It also explains slower hiring, stable margins, and delivery velocity gains without any visible shift in contracts.

Digging Deeper

Meanwhile, the middle layer makes AI visible but not central. This includes functionalities like chatbots in a migration programme, generating automated summaries or using agentic assistance.

“These aren’t new deals. They’re existing scopes with AI components added inside,” Gogia says. “They still follow traditional pricing and still get delivered by core teams, and often don’t require any change in how the client thinks about governance. AI is there, but it’s living inside old wrappers.”

Vasu sees the same pattern across most client programmes with internal copilots, process automation for invoices and claims, code modernisation using generative AI and customer chatbots, and agent assist.

“Very few projects involve training foundation models from scratch,” he says. “Most use pre-trained models plus enterprise data plus workflows.”

The break begins only at the top layer.

“Here, AI is the core product,” Gogia says. “You’re building or operating AI agents. You’re delivering enterprise copilots with retrieval logic and document grounding. You’re running orchestration layers where decisions are triggered autonomously.”

This is where delivery stops being deterministic. Outputs become probabilistic. Risk moves from system uptime to model behaviour. Governance becomes part of the service itself.

“These programmes require model governance, audit readiness, risk assessment, and operational guardrails,” he says. “That’s not a bonus feature. That’s a core service with serious consequences.”

This is also where pricing, accountability, and risk sharing begin to shift. “It’s no longer just about effort hours,” Gogia says. “It’s about whether the system behaves reliably and safely.”

Yet this layer remains small. Infosys says it has 4,600 AI projects. UnearthInsight’s estimates suggest only 15–20% of those are standalone AI mandates with clear budgets and board visibility. “50–60% are AI embedded inside existing cloud, data, or app contracts,” Vasu says. “20–30% are pilots or early-stage experiments.”

“Numbers like 4,600 AI projects or 500 agents make for great headlines but need heavy context,” Gogia says. “These aren’t thousands of large-scale, board-visible transformations. They’re a mix of pilots, embedded workstreams, internal tool usage, and production systems.”

One transformation programme rolled out across geographies can become twelve projects on a dashboard. A testing team using AI once a week can become an “AI initiative.” Internal enablement often inflates the count.

“So yes, the volume is real,” he says. “But it doesn’t mean thousands of standalone AI initiatives are live across hundreds of clients.

Only a Fraction of Work is Real AI

Neeti Sharma of TeamLease Digital gives a similar assessment. “About 15–25% of AI projects are likely to be core, standalone AI programmes with business-critical impact,” she says. “The majority are AI components embedded inside existing transformation contracts.”

Topline confirms the pattern. In most large deals, AI rarely appears as a line item. It hides inside delivery assumptions and productivity gains.

“Clients are not paying a premium just for AI,” Gogia says. “They’re paying for results, and AI is how those outcomes are delivered.”

Vasu estimates that in a typical transformation deal, only 5–10% of revenue is directly attributable to AI work, while 20–30% of the delivery effort uses AI tools internally.

“Pricing is still FTE-anchored, with early movement toward outcome-based pricing,” he says. “Vendors are still absorbing most AI productivity gains rather than fully monetising them.”

Sharma puts it more bluntly. “About 30% of software development is getting done using AI, which translates to 5–10% of overall deal value,” she says. “Most revenue still comes from application, cloud, and infrastructure work.”

The bigger impact today is margin. “Vendors win deals because they can deliver faster or with fewer people,” Gogia says. “AI is central to that capability.”

So has Indian IT changed?

Yes, in how it delivers. No, in what most of it delivers.