A C-SUITE VIEW ON AIWe speak with leaders about how their companies are putting artificial intelligence to work. In this edition, Tech in Asia talks to Raunak Mehta, co-founder and CEO of Igloo, a Singapore-based insurtech firm that provides digital infrastructure across the insurance value chain. The interview has been edited for brevity and clarity.  You've spoken publicly about Igloo becoming AI-native. What does that actually mean day-to-day? Being AI-native has two layers for us. The first is across our operating model. Wherever we play a role in the insurance value chain - be it customer acquisition, policy distribution, underwriting support, or claims management - AI is playing the majority role. The second layer, which is where I think we differentiate ourselves, is how AI is deployed internally. We're an engineering-led company, so the question becomes: how are our engineers, designers, product managers, and DevOps teams using AI to do their jobs faster? The same cuts across finance, marketing, and HR. At the volumes we operate at - we peak at issuing 700 to 800 policies per second - it's humanly impossible to manage things in a manual or even semi-automated way. That's fundamentally where the need for AI came from. Where have you seen AI move the needle most for customers? Two areas. On the sales side, we use AI to make real-time recommendations based on customer signals. For travel insurance in Indonesia, we look at where someone is going and why - business trip, beach holiday, conference - and suggest relevant coverage. Eight out of 10 times, the product we recommend is the one the customer picks. We also recently launched a travel insurance sales bot called Igi, which handles everything from product discovery to comparison to purchase. In the first 10 days, 20% more users reached payment compared to those who weren't using the bot. On claims, we have agentic workflows that assess the completeness and accuracy of submitted documents within seconds. Turnaround time at the 95th percentile used to be measured in days - it's now measured in minutes. Both our retrieval and contextual accuracy rates are above 90%. Your AI recommendation engine suggests coverage upgrades based on location and other real-time signals. How do you make sure it recommends what's best for the customer, not just what drives conversions? That goes back to how we've built our AI stack. We have a foundation layer, an agentic layer - both fixed workflow systems and fully autonomous agents - and then, most importantly, a data layer. With 100 million policies a month, the proprietary data we hold across dimensions is substantial, and that's what the AI operates on. On top of that, we have an agent evaluation layer and an observability layer that provide a feedback loop, evaluating input signals in real time and adjusting agent behavior accordingly. It's not a static system. Ultimately, the success metric in a direct-to-consumer setup is conversion. We've seen that go up week over week, month over month. So the two aren't actually in conflict. Many enterprise AI deployments fail to deliver meaningful results. Why do you think that is, and has Igloo experienced its own version of that journey? I don't think AI is at fault. The biggest problem whenever enterprise AI fails is largely how siloed the data is. Large organizations run finance on NetSuite, CRM on Salesforce - systems that never talk to each other. The moment you try to move AI from prototype to production, it breaks. In our case, it's been difficult to point to moments where AI hasn't delivered. I'd attribute that to three things: the architecture, the way we've managed data over the last eight to 10 years and made it available as a knowledge base for our AI layer, and our engineering team. AI didn't start for us in late 2022. After AlphaGo in 2017, we were already exploring TensorFlow and neural networks for insurance - probably one of very few companies in Southeast Asia doing that at the time, outside of Gojek. You use frontier models but seem to have a specific philosophy around how much to lean on them. What is it? LLM plus harness. The LLM is the brain. The harness is the helmet you make it wear so it doesn't go out of its way and hurt itself. If you're heavily dependent on a frontier model for everything, you're setting yourself up for failure. Frontier models are probabilistic by nature, and in an enterprise setup, you need answers that are as close to deterministic as possible. Context anxiety, context rot - these are real problems when long conversations make outputs increasingly unpredictable. We try to avoid calling frontier models for things that could be handled through an MCP server or a simple utility function. We may also look at building domain-specific small language models for areas like motor claims, where the problem space is closed and the answers need to be consistent. Beyond that, what I'd want from providers - besides lower token costs - is more pre-built scaffolding. Anthropic's recent agent releases are a step in that direction. Right now, we're building a lot from scratch. It's a bit like writing everything by hand before libraries existed. |
