This conversation has been edited for length, clarity and continuity. A longer version is available here.

What does it mean to be an AI-first company? 

Tarafdar: The first part is: How do we use AI to amplify the potential of all the humans? The idea for us was AI is a tool that we can use to become a lot more efficient, productive and more client relevant, and become better problem finders, solvers and all. 

The second part of it is how do you weave AI into the regular ways of working at the company? If we are doing software engineering for our customers, how do we use AI to generate a lot of code? Today, if you look at it on an average, every few weeks, we generate about a million lines of code through AI. How do we make AI integral to the services that we offer to our clients?

The third part is: How do we use these to drive value for businesses? If somebody is a bank, how do we use AI in order to make their customer onboarding process a lot better? How do we make credit decisioning better? If you are a services firm, how do we use this to improve services? If you're a retailer, how do we use this to drive better customer engagement? We then start thinking about it from an industry perspective. 

The fourth is about doing frugal innovations to bring the value of AI. Eventually for businesses, it is about doing it in a trusted and secure manner, doing it at low cost and doing it with very high efficiency. We ended up building our own small language models a few months back, and we are focused on things that will drive value to the enterprise. 

What has been the biggest challenge that you have encountered turning a workplace into an AI-first workplace?

The first challenge will be the data. A lot of times, the data that is there is not fully ready to be consumed for either pre-training or building these AI solutions. We end up spending typically 60%, 70% of our time in preparing data. A lot of times the data may not exist, in which case we may have to create synthetic data in order to fix the data gaps. We know how to fix it, but it takes time. 

The second part of it is responsible AI because most businesses are regulated industries. Ensuring that we are building AI products that are legally compliant, trusted, secure, there is no bias that is explainable [or] auditable, all those things become important. Sometimes organizations may not be fully ready, either with the processes or the tooling or the risk mitigation strategies to deal with it. That takes a good amount of time. At Infosys, we have launched a solution to help organizations become faster, but we took some time as well. For us, it took about two years to get all of it in place. 

The third part of it is the cost of running AI. While the cost has come down over the last two years, it is still significant. It is not as cheap as a normal search application or a transactional application. We are looking at how do we run this with optimal cost, which is where a lot of these innovations are important: If I have to scale it to thousands of users across the company and to their end users, then it also has to be frugal enough that the ROI is justified. 

The fourth is talent. If an organization wants to build your own models, you need AI masters. Today, there are only a handful of AI masters who can build models [from the] ground up. Finding that talent is also a little challenging. Most organizations will have to either build talent or hire that kind of talent to do those kinds of activities.

What kind of advice would you give to a CIO at a company that is trying to build its own AI-first program?

One, look at the enterprise AI as a strategic driver, because this has already become a general purpose technology, which means it will get embedded into every part of the business. If that is the case, then how do we look at it strategically? 

For this, I think there are five key things that they need to get right. First is how do they find value in AI? That’s the first important thing to create a business case. For that, they need to look at strategic business value chains and not use cases. Identify areas where value can be delivered so they can demonstrate the business outcomes, which becomes critical for the success of any AI initiatives. 

The second part is set the foundation. You have the data for AI, have the platforms in place, make sure that these systems are talking to each other, the data issues are sorted. If they don’t have the right foundation, they can never scale the AI initiative.

The third is invest in the right operating model. What we have seen is just giving AI tools does not lead to better productivity efficiency or change. You also need to change ways of working. Having a talent reskilling program or the right AI talent with new ways of working is important to get this right. 

The fourth is to be responsible by design upfront. This cannot be an afterthought, because in most regulated industries, this will come to bite. That needs to be very clearly defined and done. 

And the fifth thing, which is very critical, is to have a foundry and a factory model to scale. Eventually, if you’ve got all of these right, then it is about doing hundreds of AI projects at scale. You need to have the right operating structure to do these programs at scale and democratize AI across the company, so that you have more and more innovators who are using the technology to innovate for their business and customers.