Almost every deal today has some form of AI embedded: Infosys CTO

In an interview with businessline, Mohammed Rafee Tarafdar, Chief Technology Officer, Infosys, shared insights into the company’s evolving approach to AI, emphasising how the landscape has shifted from large-scale, generalised models to more specialised, domain-specific AI solutions. Some takeaways from the interview:

If Agentic AI provides autonomous decision-making without human intervention, will we see a scaling of companies without adding much manpower?

There will be phases with Agentic AI. With AI assistants, we are largely doing augmentation, which means the same work using these AI tools for higher productivity and efficiency. The next step beyond that will mostly be automating a few tasks, either through bots or RPA tools. However, the autonomous state of agents is the last level of evolution. It will take some time before we get there because many steps are required in between. Plus, initially much adoption of these technologies will require a human in the loop because the accuracy will improve over time. Getting to a level of maturity to move to an autonomous mode will take time. In the next three years, some tasks will be automated, but we will also need human intervention to improve and iterate.

Infosys recently came out with SLMs when LLMs were the rage. What was the intention? Is the cost why some large Indian IT services companies are hesitant to invest in LLMs?

There are different roles for LLMs and SLMs. We wanted to ensure the SLMs we built are specialised to a business or a domain, that they use a lot of permissive data, and that the run cost is less. For that particular domain, it can operate at the same level or better than LLMs. That was the rationale. We were also strategic in choosing the areas. For example, we picked banking because whatever model we use here, we are integrating into a Finacle product. So, it’s a vertically integrated AI solution offered as part of Finacle. Second, we did for IT operations. We are integrating this into the LEAP platform, which runs operations for our clients, to improve productivity and efficiency.

Another is cybersecurity, which is getting integrated into our Cyber Next platform to run the SoC operations for our clients. We picked areas where we have platforms so we can integrate and create value for those domains. SLMs and LLMs will coexist in most enterprises because when you need larger generalised knowledge, you need reasoning capabilities for which you must rely on an LLM. But when you are doing specialised, domain-intensive tasks or for a business area, you want to do it in a secure, compliant manner with IP you retain, where SLMs will have a role. The cost proposition is also better than that of LLMs.

We realised the time and cost to build these larger models has come down. You can see with DeepSeek that you can build at a fraction of the cost and time, but there has to be a business reason. We found an opportunity to build it in a domain-specific manner that aligned with our IP strategy. Tomorrow, if we find value in midsize models, we will look at it. But we are seeing it more from a value and business perspective. I don’t think cost and time are a big factor, given both have come down.

If LLMs are at the core of AI agents, what foundational models are you using?

We use over 10 different models. It’s a combination of commercially available models like GPT, Gemini and Cloud, open-source models like Llama, and Mistral, and our specialised models-- our fine-tuned SLM. Our strategy is Poly AI. This means we want to pick the best model, the best AI provider and the best platform depending on the task.

What percentage of deals in your pipeline are AI or GenAI related?

Almost every deal today has some form of AI embedded, even large deals because our clients expect us to use GenAI to deliver value efficiently.