AI is increasingly finding practical applications across the life insurance value chain, from customer service and renewals to onboarding, underwriting, risk management and sales support. At the same time, insurers have to navigate challenges around scale, data privacy, governance, security and legacy technology.
In this exclusive interview, Ganessan Soundiram, Chief Technology Officer, ICICI Prudential Life Insurance, discusses where AI is seeing the most adoption in life insurance, the challenges of deploying it across complex processes, the importance of guardrails and human oversight, and why conversational interfaces could become an important way for customers, agents and employees to interact with insurance systems.
AI is everywhere, but where is it delivering the biggest business impact for life insurance today?
I think AI has been in existence for the last several years, but from the business impact perspective, we are seeing a huge amount of adoption in customer service. A conversational interface, like a chatbot or WhatsApp bot, is a digital channel that is available 24X7 and can handle customer queries. We are also seeing significant adoption in renewal-related outreach, onboarding, underwriting, and risk and fraud management. For onboarding, we do anywhere between 55% and 60% as complete digital KYC, while AI helps with Vision AI, OCR and LLM-based OCR for the remaining requirements. In underwriting, it helps with video-based medical assessment and summarising personal, professional and medical details. In risk and fraud management, machine-learning models and AI-driven solutions help flag risks during onboarding.
AI also helps the sales team with product-related features, recommendations and process-related queries through a conversational interface. To sum this up, I think conversational interface is one of the huge trends. A customer, an agent, or somebody in operations or underwriting can ask any kind of query about the process or product-related details, and that’s where I think it comes handy. We are also seeing a huge amount of adoption in the tech space, development, testing and operations, with GenAI-assisted coding, testing and IT operations. We are seeing roughly about 15%–20% of saving in the IT cost.
What has been the toughest technology challenge in modernising life insurance?
For example, if you want to bring an AI solution for a call centre, let’s say. A call centre will have a typical, if you understand the call type, subtype, it can range anywhere between 200, 300 or maybe beyond that also. So generally having AI interfacing with customers and then understanding the nuances of each and every process, I think we will have difficulty in adopting the AI. However, if it is specific-related things, for example, reaching out at a scale for one particular use case and which will have all the objection handling, that is the place where we can use it pretty easily. So we will need to understand what is the volume and what is the intent of that particular deployment. And if there are few things, few considerations, definitely AI can play a role. If it is a broad-based, minimal number of things happening on a daily basis, the deployment of AI may not be very fruitful.
When you are adopting AI, how are you balancing innovation with trust and regulatory compliance
See, the AI per se, which we have been implementing or adopting for the last several years, there can be guardrails and it doesn’t go anything beyond your company aspect and all. But if you are having LLM as the main interface, then you will have to put proper guardrails so that it doesn’t really go beyond your company-related data.
If it is ad hoc-related things, then it has to go through a man-in-the-middle kind of approach. For example, anything which can be done as a templatised one where we don’t see any issue in terms of adopting AI, where we need to do an ad hoc kind of thing where AI plays a role in terms of framing the things, in terms of helping the things, a man-in-the-middle approach where the final decision will be taken by the individuals. For example, I spoke about the models and all where for the risk management, AI/ML kind of driven thing. There, model governance is done in a regular way so as to understand what is the assumption, what is the actual, what is the expected behaviour. This particular thing is reviewed once in a month so that we don’t go anything beyond our boundary. That there should be logging happening, there should be a kind of auditing, a specific thing has to be done for the AI-led implementation.
How are you addressing data privacy and customer trust while using AI?
If you have any particular case studies that you can mention also, see in terms of data security, whatever we are following for normal, as a manufacturer, we interface with hundreds of distributors, and data security is an extremely important area for us. It’s not only with respect to AI, but generally security is of paramount importance for us.
With respect to data privacy, the rules and regulations from IRDA take care of the disclosure requirement, the incident disclosure requirements as well as what needs to be taken care of when we deal with data transfer and data movement also. So from that perspective, I don’t think there’s any significant difference between how we deploy the solution for general integration needs as well as for AI needs. However, from the data privacy and data protection perspective, there are again rules and regulations available. That’s what we follow here. For example, the consent-related things, giving the complete visibility to the customer in terms of what we are collecting and how we are collecting, giving them the choice of when, for what purpose we are reaching out to them, those kinds of details were there as part of DPDP policy.
Is there any particular strategy that you have in place as per the DPDP Act or the policy that is in place?
See, the DPDP Act clearly talks about the purpose for which we are taking the data, and it also allows the customer to keep reviewing as and when required and to keep revoking the consent if it’s required. And also how we take care of the lifecycle management of data, which is with respect to beyond, if the purpose is taken care of, if the servicing of that particular policy or contract is done, then what kind of policy that we follow, I think those details are given. And in terms of the security aspect, we follow the defence-in-depth approach, where it’s a multi-layered security that we have. And whether the data is at rest or data is in transit, it’s all encrypted into it. So there is no risk of anybody intercepting and decoding the data and all.
If you had to bet on one technology that would redefine the life insurance sector over the next few years, what would it be and why?
This is something like we also have been thinking through in terms of what to bet and how to look at our architecture as well. My view would be to look at conversational interface as a trend that we are looking at, and we have to have everything supporting that particular trend. What I mean by it here is the way currently all of us have a way to access or to demand the service from any industry or any organisation. From a kind of interface-driven approach, there is a huge tendency towards a conversational way of getting things done. For example, I don’t really do what I did earlier. If I want to get information, I would do Google.
So it’s basically not like the way I used to access the information, the way I used to do the interface engagement and all. So I am seeing a kind of trend towards the conversational interface. Whether it is asking a query about a set of data, whether it is to drive insight about anything that we have. For example, let’s say, how do you understand what’s happening in the organisation? You may ask for a dashboard, you may ask for a report, and you may ask for some kind of insights. Those are all the ways that that was happening and each and everything would happen, a set of development, a set of rolling out and all. But currently the tendency or the way we are seeing here is you can simply ask a question about what is the outlier happening in a region, particular channel, what used to be like a kind of predetermined and then evolving over a period of time can be a simple query on the data set that you are having.
Similarly, when you look at the interfaces, for example, you can say, “I am very new to the organisation as a sales guy. I am very new to the organisation. Here is my customer wanting this. What should I do?” It can be a simple query that you can place to the interface and the interface can come back and say that there are three, four things that you can do, here is a link you can access, here is something you can propose to the customer. Or you can say, “I’m meeting a customer ID for this. What can I do?” So the customer data can be looked at from various perspectives. It can come back and give a kind of brochure and it can come back with an engagement plan to plan with the customer. All of this are happening through a kind of, we are seeing a shift towards talking to the interface, typing in the interface, and then getting the required things in an easily serviceable way. This is extremely important for an organisation like us where the process is a bit more, it’s a lengthy process, and the product features may have to be explained again and again so that the customer can understand. And we also have a sales team where the attrition is more than 10%–20%. So, this particular thing comes handy.
So, in order to have this particular thing as a prominent interface that can engage with sales, with customers, with operations, with an underwriting team, you have to have a kind of backend infrastructure that can support this. So to support the conversation as an interface layer, what do you need to do? You need to have something like an agentic layer. You need to have something like a Model Context Protocol. What do we mean by that here is, you ask a query to large language models. The LLM understands the intent of the conversation, then it looks at in terms of what is the organisational capability in servicing that particular requirement, then it interfaces and then connects to the API and then gets the detail out. So, you have to have an architecture which can support the interfacing layer like conversational and then the backend supporting multi-agent or MCP or RAG. We are investing our energy and effort and costs also on this.
Finally, is there anything that you would like to address to your peers from the industry, maybe a concluding remark or something?
To the peers, I think, see, we do meet as a CTO forum and then we also engage with regulators as well as the IAB organisation. What we keep looking at here is a lot of limitations and restrictions with respect to our core admin or our policy admin system that we have currently, which is what is primarily used across, at least in our country. There are a couple of things that are dominating. Are there any active engagement in terms of modernising the core platform is something which we have been discussing, and that’s, I think, an area where I would leave it for the players to consider and think through it.