Infosys
is one of the world’s biggest IT services companies, and today it’s increasingly trying to develop and sell
AI solutions
to the world. In the company’s latest annual report, chairman Nandan Nilekani writes that as the company works towards that, “We must learn from applying AI to ourselves, be it in creating an AI-first enterprise or in accelerating the massive talent amplification that’s now needed.”
So, how much has it applied AI to itself? How much of an AI-first company is Infosys today? We spoke to chief technology officer (CTO) Mohammed Rafee Tarafdar to understand this.
Rafee says they started the AIfirst journey about 18 months ago, targeting seven work streams. Here we look at the progress and impact in each of those streams.
Software engg, migration, modernisation
Infosys developers today use an offthe-shelf code assistant. Rafee says the company is one of the largest users of Microsoft Copilot. In addition, they developed 12 other code assistants based on specialised AI models. “In areas where nothing was available out of the box, we took largely open-source models and we fine-tuned these using the code and data we had. For example, we built a specialised model for our Finacle banking product, where, let’s say we have a UI design, then looking at the UI design, it can automatically generate code end-to-end. We created fine-tuned models for Oracle, SAP implementation, for Cobol modernisation,” he says.
In one case, they reduced the model footprint so much that developers could run it using their laptop CPU instead of on expensive cloud resources. Close to five million lines of code have been generated to date using code assistants.
Rafee says the productivity benefit in coding and testing varies between 10 and 30%. “We have also seen it improves code quality sig nificantly by improving testability, reducing vulnerabilities, helping reduce a lot of tech debt,” he says.
Internal experience & processes
This involves building AI assistants for different functions – sales or customer facing teams, HR, etc – by pooling the necessary knowledge base and enabling questions to be asked. Sales teams can use these assistants to quickly find the relevant information, such as case studies or past work, to respond to customers. In HR, if someone wants to know about a policy, they can quickly look up that information, instead of raising a ticket, and waiting for somebody to respond. “Since we operate globally across different teams, this was a big challenge because you need to know the right person to reach out to. Now you don’t have to,” Rafee says.
And these are not independent chatbots; the assistants are embedded within the system. “If I’m working on Microsoft Word or PowerPoint, an assistant comes right within there,” he says.
Learning
Infosys’s learning platform Lex has been fully GenAI-enabled. It allows the learner to ask questions, and then it probes the learner on her understanding of the concept, and gradually helps build the fundamentals of the concept. It also teaches concepts by providing analogies from other more familiar domains.
Rafee says they have even created over 6,000 content courses using GenAI. “Creating courses used to take us weeks, now it has come down to days,” he says.
Making campuses smarter
The company is investing in autonomous tech, and that is currently being used in golf carts in the Bengaluru campus. The self-driven carts take employees from one building to another. It’s being piloted on one route now. AI is also being used to forecast demand for water and energy, in order to optimise their consumption. “We also use this technology for predicting equipment failures, like inverters,” Rafee says. The forecasts, he says, are very accurate, but since the initiatives are relatively recent, the company is yet to estimate the benefits from them.
Services offered to clients
All the services that Infosys offers its clients – like application development, support, maintenance, BPM – have been reimagined to see how AI can be used to deliver them better. One dimension of this are horizontal functions, like support and maintenance, SAP implementation, and modernization. The second is industry specific. “We have now created over 25 playbooks that already have an AI blueprint that we use to deliver services better,” Rafee says
Democratising AI
To accelerate innovation, it was seen to be necessary to democratise AI within the company. So, working with Nvidia, Infosys created its own AI cloud that anybody can use to get access to GPUs. Developers can build apps on it, test it, and launch it. And everything is automated, Rafee says, which means they can do it in days rather than in months.
Responsible AI
Responsible AI practices, Rafee says, are part of everything the company does. The company, he says, has invested in building a lot of tools so that everything that’s built is explainable, removes biases, ensures security, ensures no toxicity, and minimises hallucinations. “We are one of the few companies who got certified on the ISO 42001 certification for AI in software management,” he says.
Challenges, and how to overcome them
Among the biggest challenges in implementing AI is that regulations around responsible AI are still evolving. The regulations are also fragmented – different countries have different AI regulations, and in some countries like the US, different states have regulations. For a global company, how do you make sure that you know you are tracking every single thing? We have had to put a significant amount of effort and time in order to understand what can or cannot be done, what controls need to be put in place.
Secondly, the AI field is evolving very fast. There are new techniques and models that come up every other week. So, some of the choices you make will decide whether they become a challenge or not. At Infosys, we decided early on that we will be poly AI, which means we will not get tied to any one model. That has helped us well because as soon as a model comes, I can very quickly deploy and test it on our AI cloud, see whether it makes sense, whether the cost based on tokens or otherwise makes sense, and then I can change as required.
Thirdly, the challenge is change management. AI is very probabilistic, it is predicting. When it predicts, a lot of times, you find that for the same question, you get different responses, or initially when you roll out, the quality of AI predictions or responses is not that great. As an end user, I need to understand that this will not be very good on day one, but as I keep using it, as I keep feeding more information, it will improve over time. Educating and enabling that change is a challenge.
—Mohammed Rafee Tarafdar, CTO
