Indian IT stopped talking about AI as a future opportunity in Q1 FY27 and started reporting it as a line item — but the five biggest firms each bet on a different definition of what "AI for an IT services company" even means. TCS cut ~20,000 jobs and simultaneously began building a team of up to 8,900 forward deployed engineers (FDEs) embedded directly with clients, even as its own AI revenue growth decelerated from 28% to 13.6% quarter-on-quarter. Infosys grew AI revenue to 8.2% of total revenue while narrowing its full-year guidance. Cognizant's own research found one in four enterprises have paused AI deployments because the productivity gains never materialized. Coforge reported 86% of revenue from AI-led engineering (boosted by the Encora acquisition), and Mphasis said 63% of its $461 million in new deals were AI-led via its new Tria platform.
The real story is that nobody — including these five companies — has agreed on what "transforming Indian IT with AI" looks like yet. The labor being cut (rule-based, repetitive coding) is not the same labor being added (hybrid technology consultants who sit inside client environments). The pricing model for FDEs — engineers whose job is to reduce billable hours — is unresolved. And the scramble to label deals "AI-led" is happening faster than the industry can define what qualifies.
Last verified: 2026-07-31 · AI revenue is now a disclosed line item at all five firms · Volatile facts (revenue percentages, guidance, headcount) — re-check next quarter.
What does "AI strategy" mean for an Indian IT services company in 2026?
It means five different things at five different companies, and 92% of revenue still comes from traditional IT — AI is growing fast but the base it's cannibalizing is what pays the bills. The first thing to understand is that "AI revenue" at an IT services firm is not the same as AI revenue at OpenAI or Anthropic. It's the portion of IT services contracts that the vendor classifies as AI-adjacent: building AI pipelines, integrating frontier models, setting up data infrastructure for AI, or deploying agentic systems inside client environments. There is no standard definition — one company's "AI-led" is another company's "we used Copilot on this deal."
That definitional gap matters because it shapes every comparison that follows. When Infosys says 8.2% of revenue is AI and Mphasis says 63% of new deals are AI-led, those numbers are not measuring the same thing — Infosys is talking about total revenue mix, Mphasis about new contract value (TCV) from net-new deals, and neither is held to a disclosure standard the way, say, a SaaS company's ARR is audited. The EY India "AIdea of India: Outlook 2026" report argues this confusion is itself the signal: Indian enterprises are shifting from "GenAI pilots" to scaled adoption, with 47% now running multiple GenAI use cases and 10% scaling them across business functions — but only as they figure out what actually delivers ROI.
For a parallel on how the same deployment-scarcity dynamic is playing out in the AI model layer — where orchestrating multiple specialist models already beats betting on one frontier model — see our guide to multi-agent orchestration with Sakana Fugu Ultra in 2026.
How does TCS's 20,000 layoffs + 8,900 FDE hires make sense as one strategy?
It makes sense once you separate which labor is being cut from which is being added — the cut jobs were primarily rule-based, repetitive coding work that AI can now do, while the added jobs are for hybrid technology consultants who sit inside client environments and whose skill is understanding the client's business, not writing boilerplate code.
TCS CEO K. Krithivasan told Reuters the company would convert 1% to 1.5% of its ~593,798-person workforce into forward deployed engineers, giving a range of 5,900 to 8,900 FDEs. The role has been popularized by OpenAI and Anthropic, both of which have built their own AI service companies around this exact model — engineers who embed with a client, understand their software, data, and operating processes, then integrate frontier models into that environment. A bank wanting to use AI for customer service, for example, needs someone who understands the bank's systems before picking which model to use.
| Factor | The 20,000 laid off | The 8,900 hired (FDEs) |
|---|---|---|
| Work type | Rule-based, repetitive, standardized coding | Hybrid consulting + engineering, client-embedded |
| Skill | Code generation, ticket execution | Systems thinking, domain understanding, AI orchestration |
| Who pays | Hourly / T&M billable model | Outcome-based (still being worked out) |
| Why cut/add | AI can now do this work | AI can't yet do this — needs human context |
The tension is that TCS is hiring engineers whose explicit job is to reduce billable hours — the FDE embeds with a client to automate work, which means they're shrinking the very T&M revenue base that paid for the old model. Outcome-based pricing is the obvious answer, but as of Q1 FY27 the company hasn't figured out how to price FDE work at scale. For more on the underlying economics of AI service delivery, our enterprise AI token cost optimization guide for 2026 breaks down the unit economics that make this shift unavoidable.
Confidence label: Confirmed (layoffs, FDE team size, CEO quote) — Reported (exact relationship between layoffs and AI replacement, which TCS has not explicitly framed as causal).
Why did Infosys grow AI revenue to 8.2% while cutting its full-year guidance?
Because AI revenue is growing fast but is still too small to offset weakness in the other 92% of the business — and AI itself is cannibalizing the traditional IT work that makes up the base. Infosys reported Q1 FY27 (quarter ended June 30, 2026) AI revenue at 8.2% of total revenue, up from roughly 5.5% two quarters earlier, alongside $5,082 million in total revenue (up 2.4% YoY in constant currency) and $3.6 billion in large deal TCV with 61% net-new. But the company also narrowed its FY27 revenue growth guidance from 1.5%–3.5% to 1.5%–3.0% in constant currency, citing a terminated European contract and softer discretionary demand.
CEO Salil Parekh framed AI momentum as "rapidly converting into revenue" and pointed to strategic partnerships with "all leading AI companies," while CFO Jayesh Sanghrajka emphasized "resilient margins of 21.1%" despite the challenging environment. Chairman Nandan Nilekani, speaking at the company's 45th AGM, had separately disclosed Infosys crossed $1 billion in annualized AI services revenue and is targeting a $300–400 billion AI-first services opportunity by 2030.
The takeaway is that "AI inflation" — where AI services grow as a share of revenue but simultaneously compress the traditional services they replace — is now an explicit, disclosed dynamic at Infosys, not a future risk. The company unveiled an "AI First Value Framework" built around two prongs: capturing new AI-first demand across six value pools, and augmenting existing services with AI to expand wallet share. The bet is that the second prong slows the cannibalization long enough for the first to build a new base.
Confidence label: Confirmed (revenue %, guidance, CEO/CFO quotes, framework) — vendor-reported numbers.
Why did Cognizant disclose that one in four enterprises paused AI deployments?
Because Cognizant's own research, published June 9, 2026, found that two-thirds of senior leaders "have yet to demonstrate measurable business productivity gains from AI, and one in four have already paused or discontinued AI deployments" — and the company chose to publish this finding rather than bury it. The research, "Closing the AI Execution Gap: A $2 Billion Business Boost," surveyed 1,100 senior leaders at Global 2000 companies and 100 startups across 10 industries. It estimated the average company is leaving $4 million in discrete AI investment on the table per scrapped project, with a total unrealized value of $4.7 trillion across the G2000.
The most striking finding is that organizations with weak tech infrastructure who invest broadly in AI (product innovation, talent, governance) before shoring up compute and data foundations are 60% more likely to discontinue a deployment than those with similarly weak infrastructure but a focused fundamentals-first approach. The research concludes that "the organizations seeing the strongest results are not those spending the most or moving the fastest. They are the ones treating infrastructure as a prerequisite."
For Cognizant, publishing this is a strategic signal — it positions the company as a "build the foundation first" partner rather than a "ship pilots fast" vendor, which is exactly the positioning that distinguishes it from OpenAI/Anthropic's FDE-led model. The same "AI adoption gap" theme appeared in Microsoft's FY2026 earnings, where 30M Copilot seats didn't automatically translate into per-seat productivity gains — another sign that deployment, not access, is the bottleneck.
Confidence label: Confirmed (research exists, numbers cited, published date) — vendor-funded research with survey methodology disclosed.
How does Coforge claim 86% of revenue from AI-led engineering?
Coforge's 86% figure reflects the company's business mix classification, boosted significantly by the Encora acquisition completed earlier in FY27 — which itself was an AI engineering firm. Coforge reported Q1 FY27 revenue of $592.2 million, up 49% YoY, with Encora contributing $100.7 million in just two months of the quarter. AI-led Engineering accounted for 50.4% of revenue, Data and Integration 20.9%, Cloud 14.8%, and Intelligent Automation and BPM 7.0% each. The company's EBIT margin improved to 16.0% (ahead of FY27 guidance), with organic EBIT margin at 16.7%.
The important caveat is that Coforge's "AI-led" label includes an inorganic boost from Encora — the acquisition immediately expanded the company's AI engineering footprint and rebalanced its geography toward the Americas (61.8% of revenue). This illustrates the broader disclosure problem: because there is no uniform "AI revenue" definition, an acquisition-driven mix shift can show up as organic AI strategy in the narrative. Profit fell 15% QoQ due to integration costs, which is the downside of the acquisition route.
Confidence label: Confirmed (revenue figure, mix, Encora contribution, profit decline) — vendor-reported, audited quarterly results.
What is Mphasis's Tria platform and is 63% AI-led deals real product-market fit?
Mphasis reported Q1 FY27 net new TCV of $461 million, with 63% of that classified as AI-led, and credited much of the "AI-led" share to its newly launched Tria platform — which had only been in the market for roughly seven weeks before the quarter ended. CEO Nitin Rakesh said early market acceptance of Tria "elevated us to a platform-led AI partner for enterprise clients." Operating margin contracted 60 basis points QoQ to 14.8% on ramp-up costs (a common pattern when launching a new platform), and the AI-led share of the overall pipeline has climbed from 12% in mid-2024 to 70% today.
Tria is described as having two connected product motions: "Modernize," built around extracting structured context from legacy systems, rules, workflows, and process history to create "enterprise memory"; and "Optimize," for governing AI-influenced decisions. The pitch is explicitly about "governed decisions and measurable outcomes" rather than experimentation, which is a deliberate contrast with the "pilot → see what works" model that Cognizant's research says is failing.
The honest read is that Mphasis's 63% is likely a mix of genuine demand and classification marketing — deals that were already in the pipeline are being reclassified under the Tria platform now that it exists, which is what companies do when investors demand an "AI story." The same caveat applies to every "AI-led" number in this article. But the trend from 12% to 70% AI-led pipeline share over two years is hard to fake entirely.
Confidence label: Confirmed (Tria launch, pipeline shift, margin contraction, CEO quote) — Reported (how much of the 63% reflects genuinely new demand vs. reclassification is not independently verifiable).
The five bets compared at a glance
| Company | Q1 FY27 AI bet | Key number | Risk |
|---|---|---|---|
| TCS | Embed FDEs directly with clients (5,900–8,900) | AI revenue $2.6B annualized (+13.6% QoQ, decelerated from 28%) | Unresolved FDE pricing model; cutting billable hours you also sell |
| Infosys | Report AI as a % of revenue, grow share | 8.2% of revenue; guidance narrowed to 1.5%–3.0% | AI cannibalizing the traditional 92%; AI base still too small to offset |
| Cognizant | Publish "1 in 4 paused" research, sell the fix | 25% of G2000 paused AI deployments; $4.7T unrealized | Telling clients to slow down can cut your own bookings |
| Coforge | Buy AI capability via acquisition | 86% revenue "AI-led"; Encora = $100.7M in 2 months | Inorganic boost inflates the AI mix; integration costs |
| Mphasis | Relaunch around a platform (Tria) | 63% of $461M new deals AI-led; 12% → 70% pipeline in 2 yrs | 7 weeks of platform history; ramp-up costs compress margin |
Which Indian IT strategy will dominate two years from now?
The most durable strategy is a combination, not any single approach — embedding FDEs, building proprietary platforms, and acquiring domain-specific AI capability will likely coexist, because no one approach covers the full range of what enterprise clients need. The skills being cut (rule-based coding) and added (embedded consulting) are genuinely different, so the "TCS layoffs vs. hires" frame is misleading on its own — the real strategic question is which firms can build enough of the new skill before the old skill's revenue base erodes.
Three signals suggest the FDE model is the most defensible bet:
- OpenAI, Anthropic, and Amazon have all built their own FDE teams — validating the role from outside the IT services industry.
- The FDE skill is scarce and hard to hire at scale (TCS cites 1–1.5% of workforce, not 10–15%), so first movers have a real talent moat.
- FDE work can't be automated by the same AI it deploys — understanding a specific client's environment requires human context that frontier models don't have.
But the pricing question is the unresolved bottleneck. Traditional IT was billed T&M — billable hours. If an FDE's job is to reduce the client's billable hours, the FDE's own billing has to be outcome-based, and outcome-based pricing at IT services scale is still in early days. The whole industry is watching TCS to see what model emerges.
Two less-defensible bets: Coforge's acquisition-driven AI category is only as durable as the integration holds and the "AI-led" label survives scrutiny; Mphasis's Tria is seven weeks old — early traction is promising but not proof. Infosys's disclosure transparency is valuable but disclosure alone doesn't make the cannibalization go away.
What does this mean for you?
If you're a builder or small business owner wondering how this affects you, the takeaway is not "Indian IT is automating itself" — it's that the AI work enterprises are actually paying for in 2026 is integration work, not model work. The five companies above are buying, partnering, or hiring the people who can take a frontier model and make it work inside a specific business's existing software, data, and processes. That's the scarce skill, and it's the skill anyone doing AI for their own work or building an AI product needs to develop.
If you're evaluating AI tools or agents for your own business, Cognizant's research has the practical lesson — invest in your data and compute foundations before you invest in broad AI capabilities. Companies with weak infrastructure who went broad-first were 60% more likely to have abandoned their AI deployments. For the framework on turning AI investments into compounding revenue rather than one-off pilots, see our 2026 guide to building AI revenue loops that compound. The mechanics of how enterprise AI actually gets costed — the token economics that drive these IT services bets — are covered in our guide to AI token cost optimization for enterprises in 2026.
What is the EY India "AIdea of India 2026" SLM recommendation?
The EY India report argues Indian enterprises should build smaller, cheaper, domain-specific small language models (SLMs) rather than chase trillion-parameter frontier models — because SLMs are faster, cheaper, tailored for Indian languages, and deployable at the edge, which matters in regulated sectors like banking. EY frames this as a strategic advantage, not a retreat: no Indian IT firm has the capital, compute, or research talent density to sustain frontier model development against a handful of US and Chinese labs, so competing on frontier scale was never the fight. The fight is in building models fine-tuned for specific industries, regulatory environments, and languages — where large IT firms have client relationships and domain depth that frontier labs don't.
For more on why localized, smaller models matter for enterprise AI economics — and how running models at the edge changes the cost equation — see our guide to Google's Gemma 4 QAT, a 1GB local AI model deploying to any device in 2026.
FAQ
Q: Did TCS really lay off 20,000 people and then hire 8,900 AI engineers? A: Yes — TCS cut roughly 20,000 jobs in late 2025 (primarily rule-based, repetitive coding roles) and is now building a team of 5,900–8,900 forward deployed engineers (FDEs) who embed directly with clients to implement AI. CEO K. Krithivasan confirmed the 1%–1.5% of workforce figure to Reuters. The cut and added skills are genuinely different.
Q: What is a forward deployed engineer (FDE)? A: An FDE is a hybrid technology consultant who works directly inside a client's environment to integrate AI tools across their systems. The role was popularized by OpenAI and Anthropic and requires understanding the client's software, data, and operating processes before choosing and deploying a model. It cannot currently be done by AI alone.
Q: How much of Infosys revenue is AI? A: Infosys reported 8.2% of total revenue as AI-related in Q1 FY27 (quarter ended June 30, 2026), up from roughly 5.5% two quarters earlier. The company also crossed $1 billion in annualized AI services revenue and targets a $300–400 billion AI-first services opportunity by 2030.
Q: Why did one in four enterprises pause AI deployments? A: Cognizant's June 2026 research ("Closing the AI Execution Gap") surveyed 1,100 senior leaders and found 25% had paused or discontinued AI deployments because the productivity gains never materialized. The root cause was usually weak tech infrastructure — companies that invested broadly before shoring up compute and data foundations were 60% more likely to abandon deployments.
Q: What is Mphasis Tria? A: Mphasis Tria is an enterprise AI platform launched in Q1 FY27 with two product motions: "Modernize" (extracting structured context from legacy systems to build "enterprise memory") and "Optimize" (governing AI-influenced decisions). Mphasis attributed 63% of its $461M in Q1 new deal wins to AI-led demand funneled through Tria.
Q: Should Indian enterprises build small language models instead of using frontier models? A: The EY India "AIdea of India: Outlook 2026" report recommends a hybrid approach — use SLMs for high-volume, regulated, and localized applications where data privacy and predictable costs matter, and LLMs as reasoning engines for broader problem-solving. No Indian firm can compete on frontier model scale; the competitive edge is domain-specific, edge-deployable models.
Discussion
0 comments