00:24:04 · Amrit Santhirasenan, Sabine VanderLinden
And the underwriting agent can say, Okay, these are the most important factors in this model. So I'm going go and do some research on those factors. And so to me, it's like that human in the loop and the human in the loop side, I think sits really tightly across that because these agents aren't going to be able to do everything themselves. In the same way when I was a junior actuary, I had to go and get some help when I wasn't sure what to do. The humans are going to be there. So it's not too different from what we've done in the past just with a supercharged set of capabilities from these agents and the humans doing the appropriate guidance, judgment, conducting the orchestra, if you will. [00:24:42] S1: Well, you just mentioned something that I actually heard over the past three weeks, working with various conferences and hosting webinars, which affect not only claims, but also underwriting is the fact that we are now recruiting younger underwriters. And so we need to find a way to build those teams. So using the technology as a teammate to enable the more experienced underwriters to train easily the younger underwriter. What is your experience with this? Because our industry is suffering from a talent gap. And so new underwriters are coming to market. I've heard it from US to Europe. All insurers have told me that over the past three weeks. How do we leverage the technology to help us there?
00:25:34 · Amrit Santhirasenan
[00:25:34] S2: Yeah. I mean, you know, it's a... It's an existential problem, think, that we're going to see over the next couple of years in every sector. Right? Because, actually, these models and technology get capable at junior tasks or they get capable at senior tasks. And so you can see there's lots of studies being done that show how there's been a collapse in the junior job market. And so when I think about this, this challenge is existential, not just for these new underwriters who are coming in and but for all people working in insurance and in the economy more broadly. I think the really powerful thing that I think the that are the cutting edge technologies available now is that they can be used not just to do the work, but to explain what they're doing. So if you have an underwriting agent, an actuarial agent, you could say, tell me what this algorithm does. You don't need a rocket ship Hx model developer or actuary to explain to you. Right? You have it built in to our platform. You can copy a bit of code and say, explain this to me. For me, think this is going to go much further because one of the things that we see in particularly in insurance where the pricing, where the transformation projects that we tend to see in our market and that Hx has driven probably upward north of approaching 50 of them now is that they're very large in scale and they take time. And the moment that happens, knowledge transfer becomes such a key thing. People leave, people go on breaks, they have children. Right? Many, many things happen. They get reassigned to other projects. And so what we're going to see now is a mechanism of institutionalizing the knowledge implicit in these systems and having again an infinitely patient, infinitely scalable army of teachers to share that knowledge. I think that's going to be a critical kind of gap filler for people because, you know, the reality of the system is... Of the situation is even during busy renewal seasons, the underwriters and senior actuaries are busy enough as it is, right? So being able to go and ask someone or something for some guidance digitally, I think is a huge unlock. I think it's simultaneously causing a bit of the challenge and that people don't need to hire these people. But it's also going to be part of the solution, because it'll be able to do some of that teaching, that knowledge transfer itself.
00:27:51 · Sabine VanderLinden
[00:27:51] S1: Would like to dive into some of the capabilities. So I'm going to go a bit technical and you will tell me if I'm going right or wrong here. I read, right, on your website that you have that data ingestion library, right, to really tackle the mess, which is out there when we think about mission by leveraging large language models. Can you share some example as to how that works? How we start maybe parsing data, you know, releasing manually forward, improving the way we are dealing with underwriting and actual activities whilst looking at age cases?
00:28:30 · Amrit Santhirasenan
[00:28:30] S2: Yeah. Great question. And so you're right, technology has moved very fast. And I think our website in the next, few days will be updated again because the data ingestion library is now our ingestion agent. A very timely conversation. So I think, yeah, it's a great question. So I'll say, you know, one of the things that have been really wonderful for Hx is by focusing on the core problem of understanding the data requirements of our clients for seven years in their quest to make the most impactful pricing and underwriting decisions, we actually created a framework inside Renew for teaching agents what to do. I won't tell you that when I set up Hx in 2017 that I prefigured the rise of the LLM and GPT. There's nothing to do with that. It's just that actually data has always been very important to us. So when you build a model or you build a rater inside Renew, you have to, by design, you tell it the data that it cares about. And that data is used to be entered by a human or API, and then it's used to run the rating algorithm. The most beautiful thing about that is that you can use... You can get an ingestion agent to... If you were to build the ingestion agent, it already has all of the information inside Renew about the data that needs to be captured and what it's used for. And what we were able to do was build an agentic framework which can refer to the existing platform. It understands the data that needs to be captured so you don't have to do any extra work on saying the information you need to capture and it understands what it's used for. And so to answer your question specifically, what does that do? Well, allows you to do things that many other systems can't do, right? Because before they've even before you've even started in Renew, you have an incredibly rich set of context, which is now an important phrase in AI, right? Rich set of context for the agents to work with. That means we can do like very, very clever things, right? We can do things like dress cleansing, which in many other providers, technology suites is a paid for extra feature that's a separate thing to buy. That's just a property that we were able to build into our ingestion agent incredibly quickly in weeks. Right? We can do things like semantic analysis. So we can not just do a transposition or take data from a Word document and paste it into our model.
00:30:44 · Amrit Santhirasenan, Sabine VanderLinden
We can read whole blocks of text and work out the intent of what that text means and then fill out related fields inside a Rater Even though it's not a one to one mapping, because we have all of that extra context. And it means we can do just much more rich, much more intricate things than you'd be able to do if you just said, I care about these five fields. That's the tiny subset of the information inside our system like ours. Is super. [00:31:11] S1: So Amrit, as you are actually talking about the detail of the system, another question that comes often to me is how do we ensure that everything we do is compliant by design? When we start looking at those agents and you actually mentioned all the capabilities an underwriting team, an actual team, a successful team needs to have when they start using agentic capability. Do we ensure we avoid bias? We avoid errors, right? In a world where a discrimination exists. Now we are going to systems helping us do the underwriting. So how do we avoid that, right? How do we ensure that the feedback loop is always there, always robust to make sure that we adapt at the same time to current market conditions?
00:32:08 · Amrit Santhirasenan
[00:32:08] S2: Yeah. It's a really good question and a really tough one. And I would, you know, the sort of thing that people like Sam Altman and Mark Zuckerberg are worrying very deeply about and that we think about hugely as well. So I'll say a couple of things. You know, we've always had this challenge. You know, we've been talking about these things for a really long time before the of rise of Gen AI. And I think what Gen AI has done is it scaled the ability and velocity with which we can do this work. It has also added much more obscurity into the mechanisms by which it works. There's a whole new emerging field in machine learning called mechanistic interpretability, which is actually being able to look inside these models and try and understand how they actually work, how the brains inside these gigantic matrix modification engines are actually working. And, you know, this field is incredibly new. There's loads of really fascinating podcast apps on this. You know, again, if you're ultra nerd, recommend listening to some of them. I think, though, that closer to home, there's two parts to this. This is the problem of user experience. Right? First and foremost, right? I think it's an incredibly... Like, we're very focused on this. It's incredibly important to us that the moment an agent expresses an opinion, it is unambiguously clear that the agent has done that and it's not human. And there needs to be an incredible degree of precision and auditability over every single judgment or decision or entry or any other interaction that an agent has with any system. Because the reality of the situation is that we are not going to solve in the same way that we were never going to solve bias perfectly before. We're not going to solve it perfectly now. But if we can audit it, we can understand it, then we can put the appropriate judgment feedback loops on. We can set systems up for spot checks. We can set systems up to hoist best decisions and monitor them, analyze them. There's this concept of adversarial modeling. You can have a separate judgment and bias agent looking over the the other agents, right, and actually supervising it. And I think we're going to see a whole ecosystem of this technology emerge that sits on top of the AIs that are just doing the work. I think some companies are very massive, very AI native, very AI first.