Editor’s Notes: In this keynote at Infosys AI Day 2026, Nandan Nilekani discusses why the current AI transition is fundamentally different from previous technological shifts like the move to cloud or mobile. He describes the shift as a “fundamental root-and-branch surgery” for businesses, requiring a total overhaul of legacy systems and operating models rather than just adding a new technological layer. Nilekani emphasizes that the real challenge lies in the “deployment gap,” where the speed of AI advancement outpaces an enterprise’s ability to implement it effectively. He concludes by highlighting that while the opportunity for AI is vast, the primary challenge for firms moving forward will be execution, talent reskilling, and navigating a non-deterministic world. (Feb 17, 2026)
TRANSCRIPT:
Tech Transitions: Why This Time Is Different
NANDAN NILEKANI: Thank you and great to have you all here in these tumultuous times. Today I’ll talk about tech transitions. I have had the fortune or misfortune of being in this industry for more than 40 years and I’ve seen a lot of transitions. So I thought I’ll talk about less about that and more about why this time it’s different and what are the implications of this transition.
Now, safe harbor clause.
We have seen technology shifts for centuries — with the printing press or telegraph. But over the last 60–70 years, we have seen a much faster change: PCs, cloud, gen AI, agentic AI, and so on. So change of technology and the speed of change has been a constant for many decades now.
And each time there is a change, the way we address that change has been different. So we went from mainframes to many computers to PCs, client server, LAN, web computing, mobile, enterprise apps, big data. And each time we had to think of it in different ways — how do you think of it in terms of making it globally available through the internet, or how do you do enterprise IT? So each time there was a tech transition, it had certain implications for us. And firms like Infosys had to deal with what was new. So we are used to the fact that each time there is something different.
The Unprecedented Speed of AI Adoption
This time, the AI transition has been much faster than earlier transitions. If you look at the number of years it took to reach 1 billion users — the internet took more than 10 years, smartphones took 5 years, AI is taking a couple of years.
Now you have to realize that the AI speed is because of the first two things. Internet was already ubiquitous, smartphones were already ubiquitous. It therefore allowed people to distribute a ChatGPT or Gemini or Claude very easily. So in some sense, the speed of AI is also because of the infrastructure of the previous era.
A Fundamental Change to the Way Businesses Operate
Now what has happened this time is that this is a much more fundamental change to the way businesses will operate. So this is not just a layer of technology.
When smartphones came, we could build applications where instead of having a PC, you did it on the phone. It was like putting a front end to an existing application. When cloud came, we could do a lift and shift — you could take the app from your on-prem and move it to the cloud. So you could do a lot of things to get going.
But this time, it’s not that. This is a fundamental change in the way we do things. There’s a technology dimension and it’s all about having AI-native architecture. But there’s a whole business dimension to this. You cannot run business the old way. Businesses have to change, the customer journeys have to change. All those things have to change.
It’s a huge challenge for talent. Talent will have to deal with a world where writing code will not be the goal. It will be actually making AI work — orchestration and those kinds of things. So the jobs will change.
And the operating model — how do we make this at scale? How do you get a firm with hundreds of thousands of employees to change all the things and make it work?
“A Fundamental Root-and-Branch Surgery”
And of course, our mental models have to change. Technology is always deterministic — you said A plus B equals C. So no matter how many times you said A plus B, the answer was C. In this AI world, every time you give a prompt, you’ll probably get a different answer. Therefore, how do you deal with this non-deterministic world? And how do you make sure that what you build has the robustness, reliability and resilience of the deterministic world? That’s the challenge for everybody.
So this is “a fundamental root-and-branch surgery” of the way business is done, which is why this technology transition is so dramatically different from anything else that we have seen.
Modernization of Legacy Systems: No More Deferral
NANDAN NILEKANI: Now, one clear learning we have is modernization of legacy systems cannot be deferred anymore. What happened over the last 60, 70 years is people would not replace the legacy system, they just added to it. So if you go and look under the hood of a large enterprise, they will have mainframes from 1960, they’ll have mini computers from 1980, they’ll have LAN from 2000, they’ll have all kinds of things and all coexisting in silos.
That is over. If you really want a firm to take advantage of AI, you have to fundamentally clean this up. So this is a massive, massive cleanup job, which everybody is dealing with.
There are reasons for that. One is the financial drain. Many large companies are spending 60 to 80% of their IT spend on maintaining systems. There’s no business value out of that. They want to go from 60% or 70% maintenance and 30% new systems, to 30% or 40% maintenance and 60, 70% new systems. They want to flip the way they spend money, but they can’t do that without that fundamental cleanup they need.
Moreover, many of these systems were designed in an era before you could have online attacks and so on.
So security breaches, which you see every day, are just going up everywhere, and there are more state and non-state actors who are getting better at it using AI. So security is a huge problem for everyone. We have seen so many cases in the last few months.
And because the data is all in silos, you can’t even innovate fast. So there are fundamental structural issues today we have. The demand side is absolutely demanding modernization.
But the good news is, for the first time, because of AI, we have the tools now to do modernization fast and very quickly and in a much more economic way. So we have a huge demand, and we have the ability now to do it, and perhaps our team will talk about that.
So fundamentally, accumulated debt over decades must be paid. You no longer have the option to defer this. And this is a huge, huge requirement, and obviously, it’s a huge opportunity for us.
The Shift From Buy to Build
Now, the other thing which is there is, as AI becomes a bigger part of the spend, the balance of advantage is moving towards build rather than buy. And that is actually what is driving some of the concerns about what will happen to SaaS companies and all that — it’s because of this. Building applications has become so simple that very often, you may just build, or you may replace something that you bought with something to be built.
And so we are seeing that, and that again actually benefits folks like us, because when it comes to building — who’s going to build it for them? It’s going to be us only who will build it for them. So fundamentally, it’s good for us.
Agentic Interfaces: The New Frontier
And the other thing which is there is that our view is that foundational systems will increasingly become systems of record. But the interface will be agentic, because an agentic interface makes a lot of sense. An agentic interface allows people to produce something which is designed pro-consumer or pro-user. And an agentic interface enables you to take out the complexity and hide it behind the agent. So the agent is simple to use. It is a very simple idea.
Now, enterprises will therefore want to put agentic layers on top of all their applications, even if they leave the system of record the same. And that is something which will be a combination of bought-out agents, as well as building their own agents. Because finally, the agents have to be composable in a customer journey which is seamless — a mix of agents which are your own or from somebody else. Again, that requires orchestration and work which somebody has to do. So there’s a huge amount of work required once they go towards build rather than buy.
Now the other thing is the pace of change is something which obviously we have not seen.
We all know about the trillions of dollars being spent and all that. But even the technological change — in 2023, a frontier model had 100 billion parameters. Today, it has 1 trillion parameters. There were only 10 to 12 agent networks; now there are 60 agent networks. So this is only going to go up. In the US alone, there are at least five frontier models. In China, there are the big four or big five. So this is only going to go up.
In India, we have seen so much action, and you’ll see some big announcements this week on Indian-based sovereign models. So I think there are certain implications of this, because if I’m a businessman and I have to choose my technology, how do I make sure I don’t make the wrong choice? Something which I invest in today may have fallen behind tomorrow. Already people are facing this reality, and therefore how do you architect your technology so that you can deal with this rapid change is a very fundamental and structural need for enterprises. And again, they need help on that from somebody who has done this in 2,000 locations and understands the pros and cons of every approach.
The Deployment Gap
But the main thing is that the technology is far ahead of its deployment. Because of this race and spending billions on AGI and all that, the technology is moving faster than the ability of enterprises to deploy it. If you look at this chart, you can see that the model performance is going up, but the progress in implementing is not really keeping pace. Because implementing this is hard stuff.
Fundamentally, it’s about organizational change, business change, retraining your people, thinking about non-deterministic approaches, changing your data so it’s no longer in silos. So fundamentally we have a situation where there is a deployment gap between the power of the technology and the capacity of businesses to use it. So if you think that some better product coming along is going to solve this — nothing is going to happen, because the problem is here, not there. You get it? It’s about how fast companies can implement, so you have to look at that.
We call this the deployment gap, but this is actually a concept by Professor Clayton Christensen at Harvard, 25 years back. He called it technology overshoot — where technology gets ahead of the need. And in fact, he argues that that’s how newcomers come in, because newcomers can then launch new products that are not as sophisticated, but good enough for customers. And Satya, in his recent blog, talked about model overhang, which is the same idea. The technology will keep getting better and better because billions are going to be poured into it. There’s massive competition, but enterprise deployment is not going to go up at the same pace. And this deployment gap is what we can help to address. So again, it’s a very important point.
Talent Transformation
Now I think talent transformation is huge. It’s not that you will not need talent — you will. But it’ll shift from QA testing or traditional development. We have all kinds of new roles: AI engineers, power deployment engineers, AI leads, forensic analysts, data analysts. So fundamentally, the challenge will be how do you take your workforce and make sure that they are re-skilled and ready for the new business. And that’s really the challenge that all firms will face.
There will be roles. The way you hire will change, the way you train will change, the way you deploy will change. All of that is going to happen, and I think we’ll have sessions on that. But fundamentally there will be a need for people — they’ll just be doing different things.
Also, a lot of the talk about productivity is Greenfield. Writing Greenfield code is not a big deal. I can take a tool, give it to a kid, and he’ll generate a million lines of code. But that’s not the real world.
The Reality of Brownfield Systems
NANDAN NILEKANI: The real world is the fact that companies have trillions of dollars invested in their systems. They have technical debt, they have data silos, they don’t have documents. Somebody was telling me the other day that there are some old systems and on contract they have guys as old as me — 75-year-old guys — because nobody else knows what the hell is going on. And then when there’s a crisis to be sorted out, they’re pulled in from Phoenix or Florida or wherever they are, and they have to solve problems that nobody else knows how to solve. So we have that kind of situation out there — undocumented dependencies.
So taking brownfield systems and modernizing them is a hell of a lot more difficult than doing greenfield development. And a lot of us get biased because all the guys who talk about productivity are talking about greenfield development. And therefore, getting these large enterprise organizations’ productivity going is a much harder challenge.
The Risk of Fake Productivity
Also, AI implementation requires laser focus. The very fact that you can generate stuff means you can generate slop. In fact, five years from now, there’ll be more AI legacy systems than any other legacy system — all the kind of stuff that will have been generated — and we’ll have to clean that up as well.
And if an organization is not careful, you can have this fake productivity. Let’s say there are two guys and they are having a fight. One guy will draft an email that is one paragraph. He will give it to AI to make it into a 10-paragraph email because he wants to impress the other guy. The other guy will take the 10-paragraph email and summarize it back to one paragraph. So both have used AI, but what have we achieved? Nothing.
So how do we make sure AI is used effectively? You need usage guidelines, you need quality gates, you need explainability. How do you make sure that AI investments lead to real performance and productivity and not just some make-believe stuff? This is something which is very important.
What Still Matters: First Principle Thinking
So what still matters? First principle thinking. One of the things when we train people is they have to learn to do this without tools, because all of us learned to do this without tools. So when we got the tools, we knew how to use those tools better. But if you start by teaching them tools, then everything is a black box. It’s like the guy who never knows how to calculate because he was born with a calculator.
So first principle thinking is very important — all the more important as you think about strategic transformation of large enterprises.
Understanding Enterprise Context
Second, understanding enterprise context. Every company is different. Every company has a different legacy. Every company has different systems. Some of them have come from acquisitions. Some have come because they have five business units all buying their own version of technology — all kinds of reasons. Every company has a complex estate of systems. And context is essential to being successful at AI deployment. Each context is different, and it’s the dealing with this context that is the hard part — which is where, again, we believe we have a way of doing that.
I’ll give you an example. The self-driving cars — the first DARPA challenge was in 2004. The first time they rolled out was 2007, about 20 years back. Then everybody said, “Oh yeah, by next year we’ll have self-driving cars.” It’s 20 years later, and all we have is a few cities in America where there are self-driving cars — because the context is different. Every city is different. Every road is different. And by the time they come to Bangalore, it’ll be 2047, because dealing with Bangalore traffic will be a different level of context altogether.
So enterprise context is so important. And that is something which cannot be done by a tool. It has to be done by capturing implicit knowledge and making it explicit.
Agnostic Design
Agnostic design. What I mean here is: don’t get locked into a tool, because that tool may no longer be relevant — it may become obsolete.
NANDAN NILEKANI: It may be obsolete in two years. So how do you design for agnostic that can choose any system?
Getting the house in order — we talked about removing technical debt. We have to go make the house in order for this whole thing. And then massive change management. You’re changing organization, business structure, people. I mean, this is an unbelievable change management. So unless you have leaders who can do effective change, nothing is going to happen.
It’s also about strong collaboration across firms. The firms have to, because all the knowledge is implicit in the heads of different people. How do you make that explicit in one customer journey? Focus on productivity. It’s not about using AI tools. It’s about productivity out of those tools. Otherwise you’ll get false productivity, which leads to more complications. And then this is an engineering game. AI engineering is a whole way of doing things. And that is part of your change management and transition that you have to do. And that’s a big thing.
## The Real Question: Execution, Not Opportunity
So my view is there is no opportunity gap. If anything, the opportunity is bigger than ever before. So don’t get distracted by that. You should still ask the question — what is the firm doing to take advantage of this? What is the firm doing to transform its talent for this new world? What is the firm doing to design the services and products for this new world? What are they doing to tell the customer in a way that it resonates? What are they doing to make sure that the front-end conversations with clients are done properly?
These are all the issues. And I’m sure everybody will not execute the same way. So there is an execution risk in doing that. So it is not an opportunity risk. It’s an execution risk. You get it?
And therefore, the balance of assessment is — how do we know that each firm has the execution plan ready to get to where they have to get? Are they able to do it well? Are they able to do it with speed? Are they able to do it with scale? Are they able to do it with new mindsets? That is really the question of the day that all of you need to ask.
So I’m hoping that today you will hear from our team, and they’ll give you some reassurance that we are on the right track. Thank you very much.
MODERATOR: Thank you, Nandan. For our next session on the AI Services Opportunity, please welcome Salil Parekh, Chief Executive Officer and Managing Director.
Related Posts
