Over 50% of Cognizant’s code is now AI-assisted. Its COO says the technology was never the hard part
Plenty of enterprises can say they have run an AI pilot. Far fewer can say they have rebuilt how they operate around it. Cognizant, one of the world’s largest IT services providers, says it has made that leap in its own business, and that the lesson is less about models than about management.
Rohit Kumar, Senior Vice President and Chief Operating Officer at Cognizant, describes a company that has moved from experimentation to embedding AI across its operations and delivery. In our interview, he put the challenge bluntly. “Honestly, the technology was never the hard part,” he said. “Pilots were easy to start and as easy to stall. What moved us forward was treating AI as a change in how we operate.”
The context matters. IT services firms sit at an unusual vantage point, since their own delivery model is among the most exposed to AI and their clients are watching closely to see whether they can practise what they sell. Kumar’s answers offer a view from the inside of that experiment, covering adoption, measurable impact, governance of autonomous agents, talent and the future shape of global delivery. These are his claims about Cognizant’s own experience, and they are not independently verified.
Four moves that got AI out of the pilot trap
Asked what enabled the transition from pilots to scale, Kumar describes four changes. The first is ownership. “Business and delivery leaders own the outcomes,” he said, which means AI is not left to a central technology team to justify.
The second is foundations. Cognizant invested in “clean, connected data and common platforms like Cognizant Flowsource, which brings generative AI into each stage of software delivery.” The third is early governance: “we set guardrails early on data use, review and accountability.” The fourth is visibility. “We made AI use visible,” Kumar said. “When teams can see where AI is part of daily work, they act on it. What gets measured gets acted upon.”
He adds a people dimension that is easy to overlook. Cognizant encouraged employees to experience AI directly. “Tens of thousands of colleagues built something themselves during our Vibe Coding event, a global initiative designed to democratize innovation and build AI fluency at scale,” he said, with certifications through frontier AI model companies and hyperscalers helping build confidence.
Where the gains show up
On measurable impact, Kumar points first to the routine parts of delivery. “The clearest gains show up in the routine parts of delivery, where time used to disappear quietly,” he said. In software engineering, AI drafts code, tests and documentation alongside engineers, and “over 50% of our code now AI-assisted.” The practical effect, he says, is that “the waiting between steps shrinks, and engineers spend more of their day on design and review.”
In operations, the change is from reaction to anticipation. “AI reads patterns across tickets, system logs and alerts and flags likely problems before they become outages,” Kumar said. Workforce fulfilment, meaning matching the right person to the right project, follows the same logic, with demand signals surfacing earlier and risks flagged before they become delays.
That, in his view, is where resilience comes from: “fewer surprises and faster recovery.” He is careful not to reduce the story to speed. “The bigger benefit isn’t speed. It’s that people get time back to think, and that shows up in client outcomes more than any productivity metric.”
Managed services built backwards from outcomes
The shift is perhaps most consequential for the managed services model. “Traditional managed services were built around effort: how many people, how many tickets, how many hours,” Kumar said. “AI makes that model look dated.” As automation absorbs routine work, he argues, the real question becomes what outcome the client is paying for, whether uptime, faster resolution or a better customer experience.
Cognizant’s response is to design services from the outcome backwards. “We work out what rules-based automation does best, where an AI agent can take the first pass, and where human judgment is essential,” he said. Governance changes accordingly. “We manage agents much as we’d manage a team, with a clear scope, performance measures, audit trails and an escalation path,” he said, naming Cognizant Neuro AI Engineering as one of the platforms that helps run that lifecycle.
The balance, he stresses, is not permanent. “It shifts as trust is earned, but people always stay accountable for the outcome.”
Where autonomy stops
The sharpest part of the conversation concerns agentic AI, which acts rather than merely assists. “An assistant suggests. An agent acts,” Kumar said. “Once software can update a record, trigger a payment or reassign work, you have to be explicit about who owns each decision.”
His answer is a tiered model. Routine, reversible, high-volume tasks can run on their own with monitoring. Decisions that are costly to undo, affect people’s livelihoods or carry regulatory weight need human approval. In between, agents prepare the decision and people make the call.
Two principles anchor the model. The first is that “accountability never moves to the machine. Every agent has a named human owner, just as every process does today.” The second is a simple test for the limit of autonomy: “If an action can’t be undone, or we can’t explain why the agent took it, it stays with a person.”
Kumar also rejects the idea of human oversight as a formality. “‘Human in the loop’ shouldn’t mean a rubber stamp,” he said. “It should mean someone with the context and the authority to say no.”
Reskilling and the entry-level question
One of the most debated questions in the industry is what happens to junior roles when entry-level tasks are automated. Kumar starts with a prerequisite for reskilling at all. “Reskilling at scale starts with knowing what skills you actually have,” he said. Cognizant has been moving from job titles to a skills view of its workforce, identifying which skills are core, growing or fading, so it can “plan training and hiring against real demand rather than react to it.”
On fluency, he is sceptical of classroom approaches. “AI fluency doesn’t come from a course catalogue. It comes from using AI on your own work, with a coach nearby and room to experiment.” The company’s Synapse skilling initiative extends that approach to people entering the job market.
On careers, he offers reassurance with a caveat. “Entry-level roles remain essential, even as ways of working may change in some areas,” he said. “A new graduate today can review AI output, test it and learn the domain faster than before. The ladder is still there. The first rungs simply ask for judgment earlier, and less repetition.”
The shape of delivery to come
Looking ahead, Kumar expects the familiar delivery pyramid to give way to something flatter. “Global delivery will look less like a pyramid of people spread across time zones and more like a network of people and AI agents working on shared platforms,” he said. “Location will matter less than skills, context and speed.” Managed services will increasingly be judged on outcomes, and “the line between building a system and running it will blur, because AI makes it possible to keep improving a system while it runs.”
His advice to enterprise leaders is notable for what it leaves out. “The most important shift to prepare for isn’t a technology one. It’s how they plan,” he said. “Most organisations still plan in headcount: how many people, in which roles. The better question is what capacity they need, and what mix of people and AI delivers it.”
That means, he said, “hiring to a plan rather than at the last minute, building skills ahead of demand and redesigning roles before the tools force the issue.”
Kumar’s answers converge on a consistent theme: scaling AI is an operating model problem before it is a technology problem. Clear ownership, connected data, early guardrails and visible measurement come first. Agents are then governed like team members, with named owners and defined limits. Planning shifts from headcount to capacity.