# The Future of AI Insurance : An interview with Karthik Ramakrishnan, the CEO & Co- Founder of Armilla Assurance

- **Show:** The Leadership in Insurance Podcast
- **Published:** 2 Oct 2023
- **Source:** https://shows.acast.com/theleadershipininsurancepodcast/episodes/the-future-of-ai-insurance-an-interview-with-karthik-ramakri
- **Transcript:** https://podcasts.opennash.com/shows/leadership-in-insurance/the-future-of-ai-insurance-an-interview-with-karthik-ramakrishnan-the-ceo-co-fou-f639b5/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

AI models are probabilistic, meaning they always carry some chance of error, and that error risk is new enough that existing insurance does not price it properly. Karthik Ramakrishnan argues that AI risk must first be measured, through independent assessment and ongoing monitoring, before it can be transferred through a guarantee backed by insurers. Armilla offers this for third-party AI models.

The central tension is between AI's potential scale and the trust deficit that slows it. Karthik says adoption has barely begun, yet one flawed model can repeat the same mistake across every case, as in the Dutch tax scandal. He argues the answer is neither blind enthusiasm nor caution, but measuring risk, detecting failures fast, and having protection when things go wrong.

## Key insights

1. **AI risk must be measured before it is priced** Karthik says cyber, errors and omissions and general liability policies cover AI only partly and do not price its risks properly. Armilla therefore quantifies the risk first, then offers an insurance-backed guarantee.
2. **One AI model can multiply a single error** Karthik contrasts 100 human claim reviewers, where two mistakes affect about 2% of cases, with one model repeating the same flaw across all cases. This scale effect is why failures are more dangerous.
3. **Models keep drifting after they are deployed** Data drift is when incoming real-world data departs from the training data. Concept drift is when what the model has learned shifts. Both mean a vendor model needs ongoing monitoring, not just a one-off check.
4. **Customization means there is no single AI score** A base model like ChatGPT can be scored generically, but retraining on a company's own data changes its behavior. Armilla therefore gives a baseline score and checks the customized version stays within acceptable bounds.
5. **Lloyd's Lab turned an idea into a product** Armilla entered the three-month accelerator with an idea and left with its first product, a performance guarantee. Mentors from Greenlight Re and Chaucer then provided capacity, alongside Swiss Re.

## Questions this episode answers

### What does Armilla Assurance actually do?

It tests and verifies the performance of AI models that companies buy from third parties, then guarantees that performance. If the model degrades on the agreed measure, Armilla reimburses the company to protect its return on investment.

### Why can't my existing cyber or liability insurance cover AI risks?

Karthik says those policies may technically respond, but they do not adequately price in the new risk, because AI risk has not yet been properly measured. He argues the risk must be quantified first so insurers can price it.

### What is model drift and why should a business care?

Drift is when an AI model's inputs or learned behavior gradually move away from what it was built for, so its predictions become less reliable. Karthik says it is not bad in itself, but it must be monitored to stay within acceptable limits.

### What went wrong in the Dutch government AI example?

Karthik says a tax department model meant to detect fraudulent child welfare claims got it wrong and targeted poorer households, and the government had to resign. He uses it to show how AI errors hit many people at once.

### Why is a more accurate AI not always the better choice?

Karthik says the most accurate model may still expose a company to a very severe kind of risk in a particular use case. Companies should weigh the size and impact of the uncertainty, not just the accuracy figure.

## Facts

- **Karthik Ramakrishnan** [0:00:47]: "what we do at Armilla is essentially verify and guarantee the performance of AI models."
- **Karthik Ramakrishnan** [0:01:30]: "should the performance of the model degrade and whatever KPI that we are assuring degrade, we will reimburse the enterprise as a protection of their return on investment."
- **Karthik Ramakrishnan** [0:08:19]: "And 55 % of the issues that they see with AI come from these third-party models, right?"
- **Karthik Ramakrishnan** [0:09:59]: "we're moving from a world of deterministic decision-making to probabilistic decision-making."
- **Karthik Ramakrishnan** [0:12:50]: "So data drift means the incoming data that's now in the wild may be continuously deferring from the data that the model was trained on."
- **Karthik Ramakrishnan** [0:13:32]: "concept drift is that the the concept that the model has learned is also drifting from the baseline"
- **Karthik Ramakrishnan** [0:15:09]: "you've had issues where this turned into an issue, I think, for Samsung or one of the companies."
- **Karthik Ramakrishnan** [0:23:13]: "the majority of Dutch government had to resign because of an AI model."
- **Karthik Ramakrishnan** [0:23:13]: "This was a model that was used in the tax department that was detecting fraudulent claims."
- **Karthik Ramakrishnan** [0:23:47]: "it ended up targeting sort of socioeconomically challenged segment of the population."
- **Karthik Ramakrishnan** [0:18:16]: "humans in one of the plants that we were talking to these they think they have an accuracy rate of 96"
- **Karthik Ramakrishnan** [0:27:38]: "it's not a matter of if it will fail, it's a matter of when."
- **Karthik Ramakrishnan** [0:29:04]: "this is a 400 or 4 ,000 year old solution, insurance is not new to a brand new technology."
- **Karthik Ramakrishnan** [0:31:22]: "My co-founder and I have been in the AI space for the last 15 years."
- **Karthik Ramakrishnan** [0:32:00]: "even though my wife is a chief compliance officer at a large insurer here in Canada, my understanding was superficial."
- **Karthik Ramakrishnan** [0:33:26]: "So we spent three months presenting the problem, getting their feedback and really workshopping, is this real?"
- **Karthik Ramakrishnan** [0:33:26]: "And we exited with creating our first product, which we launched a couple of weeks ago, which is the product guarantee."
- **Karthik Ramakrishnan** [0:33:26]: "in partnership with Swiss Re, Greenlight Re, and Chaucer, the Lloyd syndicates, who really innovative companies"
- **Karthik Ramakrishnan** [0:35:02]: "So Greenlight Re and Chaucer were our mentors in the lab."
- **Karthik Ramakrishnan** [0:35:02]: "the largest one I saw was 15 or 30 trillion in economic activity that is going to get impacted by AI, right?"
- **Karthik Ramakrishnan** [0:39:18]: "Is there a mutual way of doing this, right? A shared risk pool. Is there alternative sources of capital?"
- **Alex Bond** [0:30:59]: "the Lloyd's app is where we met you for the first time um cohort 10 when you and we were there from pitch day to midpoint trade show"

## Transcript

**[0:00:04] Alex Bond:** good morning and welcome to the leadership and insurance podcast i'm your host alex bond and i am very lucky today to be joined by um karthik ceo co-founder of Armilla Assurance um we just discussed that and i'm thinking i hope i've got that right is are we just armilla or are we armilla insure or armilla assurance assurance there we go embarrassingly if anyone had heard we literally had that conversation about 10 seconds before we hit record so um it shows all all best laid plans and things um karthik before i go and massacre any more of your company name and um profile uh perhaps be great if you could introduce yourself and the armilla business excellent thank you for

**[0:00:47] Karthik Ramakrishnan:** having me uh alex um so yeah i'm karthik um ceo and co-founder of armilla assurance um and what we do at Armilla is essentially verify and guarantee the performance of AI models. With the proliferation and adoption of AI, one of the biggest challenges for enterprises is being sure about what they're procuring from third-party models. So we help enterprises break down the risks that are presented by these models with our technology and our platform, which forms our MULA verified and assurance report.

**[0:01:30] Karthik Ramakrishnan:** And then we provide a guarantee to our assessment as well. Essentially, should the performance of the model degrade and whatever KPI that we are assuring degrade, we will reimburse the enterprise as a protection of their return on investment.

**[0:01:50] Alex Bond:** Okay. Wow. um so this is a kind of like net new risk isn't it this this is this is where you know and and you know have to be transparent and be flattering you you are certainly a business that's been a lot of buzz about it's been a lot of excitement we work very closely with the Lloyd's Lab and we want to dive into that Lloyd's experience later but there was a lot of a lot of excitement from people when we said we were going to have you on the podcast because you fall into that absolutely net new risk category.

**[0:02:18] Alex Bond:** This hasn't been done before, this sort of insurance contract. Is that correct?

**[0:02:25] Karthik Ramakrishnan:** Definitely for AI models, hasn't been done before. It is absolutely an emerging category of risk. Anytime we get into, I'll stick to the software space. Anytime we get into new technology, It presents risks or risks emanate in different ways, right, which may not have been seen before, or it would be the same risks, but the source of those risks is that new technology. So with AI, there is absolutely a whole host of risks, how they emerge, how they present themselves to an organization.

**[0:03:04] Karthik Ramakrishnan:** And now the question is, how does the organization and enterprise evolve in their risk management practices and protection of the organization from those risks? And that's where now we need to have some innovative new ways of risk transferring those exposures.

**[0:03:26] Alex Bond:** it falls into that kind of category and i was talking about this um with a cyber mga um and we were talking about it falls into that category of it's not enough now just to offer insurance products that you know play at the point of failure in that we have to work with insureds or potential insurers on education evolution of kind of risk management and we were talking about cyber but we were saying even on the personal level you know uh health insurance you know i i have my smart watch on and it monitors how often i go to the gym and that impacts kind of how the price of my health insurance um in your view is is that

**[0:04:10] Alex Bond:** the only way that we're going to see insurance evolve because that's that's my belief is that we kind of we've we've been at arm's take for too long and now it's actually about giving people the tools to kind of improve risk

**[0:04:23] Karthik Ramakrishnan:** management with these emerging risks yeah absolutely i think the the uh as we digitize different aspects of you know our personal lives our corporate lives uh you know the the capability of measuring these risks is evolving, right? So this is not that the risks have changed in sort of established protections. It's how we monitor and measure those risks definitely is going to change. I mean, let's take auto insurance, for example. If we truly believe that self-driving cars will take over the market in the next decade or so, So how we measure and who is even the insured now changes and who provides the insurance also will change.

**[0:05:16] Karthik Ramakrishnan:** And so the evolution is the insurance industry itself keeping up with the evolving trends and technologies. And, of course, there's new categories of risks that are emerging as well. But now we have climate risk as a big sector that is growing. And we're trying to understand what that means. for an organization. And similarly, AI, the risk of AI with the scale of adoption that we expect AI to have, that risk is going to be pretty massive.

**[0:05:53] Karthik Ramakrishnan:** And we have just gotten started in even putting our arms around in understanding what those risks are. Some of those risks are very specific to AI models themselves. And some of these risks are existing risks, but they're impacted by this new technology. So in both these categories, we need to understand how we measure and monitor for that risk, and then, of course, provide the appropriate risk transfer. And absolutely, education is important. Risk managers are, you know, your second line of defense or your third line of defense are removed from the first, you know, the front office, right, which are implementing these solutions, improving the business from a revenue standpoint or a cost efficiency standpoint uh they need to now start

**[0:06:39] Karthik Ramakrishnan:** so they're leading the pack and so risk managers now need to keep up and understand what they need to do and so that's where the education component comes through uh quite well and then of course there's a two-way dialogue between the front and the back uh as to how do they manage that on an ongoing basis where do you i think that's a really

**[0:06:58] Alex Bond:** interesting point because i was thinking about the kind of risk manager is it is almost absent from the decision making of which ai tools or ai models that are being implemented um it doesn't where we're talking about in this context doesn't sound like a good idea so where do you come into the situation or where's the kind of best point for you to come into the situation because once the horse is bolted and someone's implemented something which isn't fit for purpose, presumably it's uninsurable?

**[0:07:32] Alex Bond:** Do they have to go through the framework? How does it work in practical terms about where you get involved?

**[0:07:37] Karthik Ramakrishnan:** Yeah, so typically where we are involved is, you're right, in that the first line, the front office or the business makes a decision on what technology they need. Risk managers are not necessarily in the decision-making process. They're consulted, for sure, but they're not the decision-maker. But once that decision is made, part of that decisioning of, and let's stick to third-party sourcing. So third-party models, there was a recent report that we contributed to between MIT and BCG, which said that more than 70 % of the companies are using third-party AI models.

**[0:08:19] Karthik Ramakrishnan:** And 55 % of the issues that they see with AI come from these third-party models, right? So when you're sourcing externally, while the business has chosen a specific vendor, part of that decision process is to consult with the risk managers. And that's where we step in. We come in and do a full assessment of the AI model, both from a qualitative and a quantitative standpoint. So things like, will the model actually deliver, does the model actually deliver on its stated goals?

**[0:08:58] Karthik Ramakrishnan:** And the stated goal could be a level of accuracy or a lack of bias in the model. Or if it's generative AI, the fact that it has not been trained on copyrighted or proprietary data. So there's many multiple dimensions of this risk. So we put a comprehensive review of the model and the product, really, the model drives the product. And in the context of risk managers, help inform them as to what their exposures are, what the severity of that exposure could be, and then make the right decision.

**[0:09:30] Karthik Ramakrishnan:** So now we need to be even more involved with, we, and I say we, I mean, the risk managers need to be even more involved in that final decision as part of the decision process to select a vendor. And once that's done, then you go and deploy that into production. So it comes before production. And then the part two of that is once it's in production, how are you ensuring that the model continues to deliver? And this is, again, very specific to AI.

**[0:09:59] Karthik Ramakrishnan:** we're moving from a world of deterministic decision-making to probabilistic decision-making. And if we know probability, there's a level of uncertainty into every outcome of a probabilistic system. And so that uncertainty now represents a level of risk. So how big is that? What's the impact of that uncertainty for that use case and for that company? and then you can make the right choices. If that's the vendor, it may be the most accurate model, but in another case, it might be exposing yourself to a very severe type of risk.

**[0:10:38] Karthik Ramakrishnan:** So you may make different decisions on that buying process. And then, of course, ongoing monitoring to ensure that that state of gold continues to be delivered.

**[0:10:48] Alex Bond:** I think I started with a sort of simplistic view of saying, if you're buying a third-party vendor, you know let's say piece of software that uses an AI model um you would have a certain understanding of that you would have a certain analysis of that but and forgive me if this is kind of a layman's question but it's more about the interaction between the business buying it and the the uniqueness of that with how they interact with that software how they use that software and how they propose to use it rather than you being able to give a score for a piece of software and then it just has a score.

**[0:11:27] Alex Bond:** It's about how that company is choosing to use that software. Is that correct?

**[0:11:32] Karthik Ramakrishnan:** Absolutely right. Now, you may have generic software. You know, let's say you were using a SaaS solution. And let's take ChatGPT as an example because everyone's familiar with ChatGPT these days. So if you're using the version that's offered by OpenAI, that's one model, right? And so, you know, you can do one sort of analysis of that. and it can give a score, right, on different metrics. But when you come to the enterprise, specifically with AI, it gets retrained or customized a little bit or a lot with the enterprise's own data, right?

**[0:12:10] Karthik Ramakrishnan:** So you will have a foundation model or a base model, which then gets trained with the additional information about that enterprise, which then gives, you know, the right outcomes that are tailored to that organization. So that's where you're seeing it's not a one single score, but there's a baseline score. And then you customize that to ensure that in that context that the model still is within the bounds of what's acceptable.

**[0:12:36] Alex Bond:** Yeah. And where does the, you mentioned the degradation, you're keeping on that. Where does that and how does that degradation show itself over the course of time?

**[0:12:50] Karthik Ramakrishnan:** Yeah. So, and this is particularly true for self-learning models. So the promise of artificial intelligence really is that it continuously improves itself as it encounters more information about the world it exists in. And that improvement is what keeps it abreast. Now, the challenge with that as well is that improvement may not be what you need. It may be veering or drifting. And that's the technical term that's used. So you have things like data drift. So data drift means the incoming data that's now in the wild may be continuously deferring from the data that the model was trained on.

**[0:13:32] Karthik Ramakrishnan:** So the model is learning. But if the model is too brittle, it may not be able to deal with that drift. And now your decisions or your predictions from the model are off. and so now that now you get a concept drift so concept drift is that the the concept that the model has learned is also drifting from the baseline some of it is so you need to monitor to make sure that that it's it's drifting in the right direction drift is not a bad thing if it's self-learning model you just need to make sure that it's within the constraints it's not wearing too far off uh the home base of of what what what you want to do

**[0:14:09] Alex Bond:** it's like you're picking up picking up bad habits and then going off in one direction that's really interesting um you wrote um a piece quite recently um called the adoption of ai and the trust deficit increasing um well i wanted to sort of extrapolate on that a little bit we'll obviously share the link to that as well but i just thought it was a it was really interesting about the trustworthiness and trust deficit and and whenever i think about the conversation about ai it goes into trust but what's interesting is something that you've already said in this conversation is that i i was going to ask you about something about what varies to widespread adoption well if 70 businesses are

**[0:14:53] Alex Bond:** seemingly using some ai already doesn't seem like widespread adoption is is being restrictive It's more how do we know what wants to trust and what not to? So, yeah, I wondered if you'd sort of dive into that piece that you wrote.

**[0:15:09] Karthik Ramakrishnan:** Yeah. You know, this is – it's not – look, if every single person within an organization has used ChatGPT, and again, I stick to that because it's top of mind for everyone, and a lot of people have experimented with it. So, you know, you've had issues where this turned into an issue, I think, for Samsung or one of the companies. But essentially their employees were going around and experimenting with ChatGPD. They inadvertently ended up putting in proprietary confidential information about the company into the model.

**[0:15:49] Karthik Ramakrishnan:** So now that's available publicly. So just on that point about adoption, we have adopted AI explicitly, but a lot implicitly. Right? So it's not that there's a, that's why you would see a large, you know, when you see the adoption curve, but within an enterprise, the scale of what AI can do, we barely touch the surface. That's why we don't say adoption is at the maximum because you can essentially infuse AI into most digital processes or even manual processes amplified by these models or agents that can transform the business end to end.

**[0:16:35] Karthik Ramakrishnan:** And we're waiting for that world. and one of the things that I like to think about is you know just as we have self-driving cars today so we've taken a manual effort of driving a car to now completely digitizing that it's completely autonomous we're going to go to that same process with the organization and I think we're going to go into a world of self-driving organizations or autonomous organizations where agents are doing a maximal number of the decisions with humans really coming in and helping shape those models and ensuring it's headed in the right direction.

**[0:17:09] Karthik Ramakrishnan:** So what that means is we're not there yet at all. And so we're barely scratching the surface of adoption of AI. So that's on that first part. And so now we've come to trustworthiness. As I said, no model is 100 % accurate just by its very nature. These models are probabilistic. There is room for error. There's an error quotient built into the very math of these models. And so that presents a risk. That's one way to look at it.

**[0:17:44] Karthik Ramakrishnan:** Another way to look at it is the data that you train a model on is only a representation of the world that it's going to operate in, right? So let's take an anomaly detection model in manufacturing, which is now my favorite one. So you have a vision model that looks at parts that are coming out of a manufacturing line, and you're seeing issues. It's supposed to detect faulty parts. Now, it's been trained on previous faulty parts.

**[0:18:16] Karthik Ramakrishnan:** But let's say there's a brand new fault that no one's seen before, and that's now come into these parts, will the machine pick up on that anomaly? And if it doesn't, there's an error. A really good model should pick up on most of these, and as an edge case, may not pick up on it. So how do you trust this model? So the plant manager, whose accountability it is to ensure the quality assurance of their product and the output of their factory now has to trust this model that it will get all of it right and today the humans in one of the plants that we were talking to these they think they have an accuracy rate of 96 even

**[0:19:05] Karthik Ramakrishnan:** humans get it wrong so if you introduce a model yes it will give you a lot of cost efficiencies you don't you know you don't need to have 10 line line you know operators looking at the faulty parts for visual inspection, a model does that, lots of cost savings. But if the model is only 90 % accurate, what happens then? So now you have an incremental issue, i.e. from 96 to 90%. And then second, how do you now trust that it will pick up on issues that maybe a human is more trained, a better trained to pick up on just because of our faculties?

**[0:19:42] Karthik Ramakrishnan:** So that's just one example of trust. So there's the inherent uncertainty in the model. It's the specific use case type issues that it emanates in. So the data that it's trained on, right, has to be representative, how well it's done. So when you have these challenges, that's where the trust deficit comes in, right? And what we will, again, I don't see a world where we will be able to 100 % trust the models. but we can get close.

**[0:20:15] Karthik Ramakrishnan:** And that's the effort right now. How close can we get to understanding what we put in, fully put our minds around it, understand the risks that this presents, where can this go wrong? And the better we do that, the closer we'll get to that objective of greater trust in these systems. And finally, safeguards. So when issues do occur and you anticipate that we do live in a, you know this probabilistic world and there's room for error how do you gracefully handle those error situations detect them identify them and um deal with them appropriately as quickly as possible right so that's the that's how we're going to get to the test deficit but

**[0:21:01] Alex Bond:** that's where it comes from i think some of this um i think some of the trustworthiness and the lack of it is sometimes I think is driven by you get the evangelist leading the charge the early adopters but they're so and this might be a broader point but I think everything has to be you have to be pro or against something you're not allowed to go I'm really pro AI but I want safeguards because you're seen as you know backward and I think the kind of the purest evangelizing about AI is a certain sense is to its deficit because some people want to unbridle, just let it run wild.

**[0:21:43] Alex Bond:** And you go, well, you can't because of things like drift. And when you explain it as you do, that the best models in the world will still make errors. But then the other side of that is true as well. If we don't build enough trust, we're not going to adopt broadly enough. We're not going to be brave enough with our adoption. And therefore, if I understand the technology at all, we probably won't be giving the model enough data.

**[0:22:09] Alex Bond:** We won't be feeding it enough to continuously learn so it gets better.

**[0:22:13] Karthik Ramakrishnan:** Yeah, I think the answer is not we won't be. I think the answer is how do we? Right? And I think if we understand those gaps, then we can figure out the solutions. And that's what we're really good at as humans, right? We figure a way out. Now, I'll take a step back as well to say, look, I mean, this is what makes us human. We have the ultimate optimists and we have the ultimate cynics and pessimists.

**[0:22:36] Karthik Ramakrishnan:** And we need to live in a world that's somewhere in the middle of that, right? And that's an ideal world. And we do need the optimists to push us forward. And we do need the more retracted ones to understand, okay, what are the dangers and ensure that we're doing it in a responsible manner, right? I'll give you an example why this is important. Right. And I also want to make sure that this, that when we talk about optimism and, you know, let it run, there's a whole domain of existential risk that people are worried about.

**[0:23:13] Karthik Ramakrishnan:** And that's where a lot of this comes from, I think. And I don't want to step into that right now. I think it's a very deep conversation. We can have a whole other episode on that. But coming back to an example, right? So most people don't know this, but two years ago or three years ago now, the majority of Dutch government had to resign because of an AI model. This was a model that was used in the tax department that was detecting fraudulent claims.

**[0:23:47] Karthik Ramakrishnan:** And this was particularly related to child welfare reimbursements or claims. on the tax sheets. And unfortunately, the model got it wrong. And it was, it basically, it ended up targeting sort of socioeconomically challenged segment of the population. And so as it is, you have economic challenges, financial challenges, and now you have the tax department coming after you saying, you need to return this benefit for whom it was designed in the first place. And so that was very embarrassing, and the government was forced to resign, right?

**[0:24:27] Karthik Ramakrishnan:** But it's not the fact that the government was forced to resign. Take a step back. Think of the number of consumers that were impacted by the use of that model. You had an optimist, probably unbridled, didn't check it, and now you have lives impacted. Right? And so we have to remember that. Who is being impacted? And when these things happen with AI, they happen at a scale that was not there before. So previously, let's say we had, I don't know, just for a number, 100 agents looking through these forums and adjudicating whether the claim is right or not.

**[0:25:03] Karthik Ramakrishnan:** Now you have one model, right? And so actually, so if you had 100 agents and two of them were not trained properly or whatever the case may be, they felt something was off and it was just two of them. But 98 % of the claims went through. But now when you switch that to one model, that same 2 % issue could be exacerbated across 100 % of the cases because you have one model perpetuating and scaling the same issue.

**[0:25:29] Karthik Ramakrishnan:** That's why this is dangerous, right? And so doing this responsibly is important. And we also have to be pragmatic in that we're not going to catch 100 % of these cases. And that's why, what do you do to protect when these issues occur? How quickly do you detect, identify, and remediate those issues is going to be critical.

**[0:25:55] Alex Bond:** It made me think about the governance of businesses more broadly as well. because I mean you know everyone has had this situation where you're dealing with a usually a big business and the something has gone wrong in the process so you know I've bored everyone to death on the podcast about I had all these problems with my I just could not get a bank card delivered to our office and everything went off this card and it just stopped for no reason and there was something going wrong but there was no way of coming out of the loop of the process so you just had to keep going back into the process and you know and it reminds me to a

**[0:26:34] Alex Bond:** certain extent of of that is that there needs to be a way of addressing anything that does go wrong that um accepts them that there is a possibility that the model has gone wrong and and and the way i the reason i'm drawing that analogy is that we kept having the same conversation and then we're going yeah but this is the process and i was like but what do you do when the process is broken because what's happened is the process isn't working so you know for the same situation if if the model is is broken down at some point or is giving a wrong wrong result it's building up a framework of going okay what do we do then

**[0:27:09] Alex Bond:** but also building in an acceptance that that is a possibility and i think sometimes some of the unhelpfulness is about people pushing models going well they couldn't possibly go wrong and of course they can like everything most things can um and and it's exacerbate problem because exactly as you said and this is not just ai by the way right of course yeah this is life right i'm talking about a process with a very well-known

**[0:27:38] Karthik Ramakrishnan:** british bank you're just like wrong and i don't know why but it is yeah exactly right i mean you can go meta philosophical about it but at the same time you know murphy's law is a real thing So, you know, it's processes, it's technology, it's people. We all have errors built into everything. And it's exacerbating when those errors are perpetuated and not rectified soon enough. And, you know, when we look at these models, I mean, software, I was having a discussion with someone, even in software, even in deterministic system, it's not a matter of if it will fail, it's a matter of when.

**[0:28:26] Karthik Ramakrishnan:** And as you keep shortening that gap, right, and the severity, then you're in a better, better position. So having said that, I think what we are trying to do is trying to ensure that we wrap our heads around as much as possible, that delta, that error rate, can we measure that and quantify it and understand the impact? And if we do, then we can step in and say, look, if it does, this is a probability of it going wrong.

**[0:29:04] Karthik Ramakrishnan:** Let us help you remediate. And this is where, this is a 400 or 4 ,000 year old solution, insurance is not new to a brand new technology. And I think insurance has stepped in as the industry and even the concept has stepped in as a means to facilitate interactions and all of the human activity, if you will, by taking some of that risk away from conducting these activities. And so all we're saying is now, okay, we know this is error prone.

**[0:29:49] Karthik Ramakrishnan:** If someone takes the right steps to ensure that that error is minimized and detected when it does occur well, then you can now get a protection. So for when things do go wrong, so that those are good risks that you can take on. So now the question becomes, how do we measure this risk? existing policies that we have in place as insurance do not adequately price this risk in. And that's the challenge, right? Most organizations, well, I've got cyber insurance, I've got my E &O, I've got my general liability insurance, et cetera.

**[0:30:23] Karthik Ramakrishnan:** It's true. It does cover, but those insurances policies do not adequately price in this new risk. and that's what we need to change. So before you can price that risk, you need to quantify that risk. So we do both where we've started quantifying that risk, helping organizations put in the right assurances and assessment practices in place. And then you can then get a suitable enough risk protection or risk transfer through an insurance product.

**[0:30:59] Alex Bond:** the the Lloyd's app is where we met you for the first time um cohort 10 when you and we were there from pitch day to midpoint trade show um I just wanted to sort of understand that experience you know what what did you come to Lloyd's kind of with and for and how much did that kind of shape where you are today?

**[0:31:22] Karthik Ramakrishnan:** Yeah, excellent question. Look, and this is with a lot of humility, I will say that we are technologists, right? That's the world we come from. But we know enough that we've also worked with insurance companies. We've built actuarial models in the past. My co-founder and I have been in the AI space for the last 15 years. So it's not something that we stumbled into in the last few. We've actually built companies around this. And so, you know, we've been exposed to how insurance steps in.

**[0:32:00] Karthik Ramakrishnan:** And that's where we started thinking from as our clients were telling us that, look, you can help us measure the risk, but the risk still exists. So now what do we do? And that's when we got to thinking that there was a gap in this market. Okay, so technologists seeing this problem, and we had a good enough understanding of this. And even though my wife is a chief compliance officer at a large insurer here in Canada, my understanding was superficial.

**[0:32:31] Karthik Ramakrishnan:** So we did a lot of research and learning through this. What we found with Lloyd's was the opportunity to get a deep dive, a crash course, if you will, on are we even thinking about this correctly? And that was a question we entered the lab with. We're seeing this problem. Our clients are seeing this problem. We think insurance is an answer. Let's go figure, workshop this through the Lloyd's Lab Accelerator. And we were lucky because maybe the timing was right with ChatGPD coming in and AI taking sort of everyone's capturing the imagination, we had several mentors from several syndicates who came in saying, yes, we think there's something here we would like to work with you on.

**[0:33:26] Karthik Ramakrishnan:** And so we spent three months presenting the problem, getting their feedback and really workshopping, is this real? And if it's real, how do we actually go about solving it? And we exited with creating our first product, which we launched a couple of weeks ago, which is the product guarantee. So we can now guarantee the performance of a model in partnership with Swiss Re, Greenlight Re, and Chaucer, the Lloyd syndicates, who really innovative companies that really understood what we were trying to solve for and agreed.

**[0:34:05] Karthik Ramakrishnan:** So that's the first foray into this. But all of that came through because of the mentorship. What we entered with as an idea and what we exited with as a solution, the idea was right. The solution got refined through that process.

**[0:34:21] Alex Bond:** Yeah Yeah. I think it's such a good example of what the lab's there for. It's to bring in new ideas, new people, new thinking to insurance. And it's to sort of allow you to kind of plug into that insurance expertise. So you go with exactly that, you know, a variation on probably what you did. And sometimes even like a failure in saying, right, this really doesn't work in the way that we're thinking about it, but it will in this.

**[0:34:53] Alex Bond:** Did they all come on board? The mentors come on board with providing capacity as well? Or was there, there was a relationship, didn't go that far?

**[0:35:02] Karthik Ramakrishnan:** Yeah, absolutely. So Greenlight Re and Chaucer were our mentors in the lab. And, yeah, I mean, that's how, you know, we got – that's how they came in to provide capacity to the product. And that's – you know, what needs to now happen is to ensure that the market understands that this is something that they need – there's an option for them to protect. and we're really excited because if AI is going to get adopted at the scale that we think it's going to get adopted, I mean, there's reports everywhere, but the largest one I saw was 15 or 30 trillion in economic activity that is going to get impacted by AI, right?

**[0:35:52] Karthik Ramakrishnan:** And so when you look at it from that standpoint, and now we know that every model has some level of risk that's just by design built into the models, 15 trillion, just do the math. I mean, that's a huge amount of risk that's going to emerge. We need to have a solution for it. And I think that was, I wouldn't say it was an epiphany, but that was certainly the realization, more visceral realization that everyone at the table, including us and our mentors got.

**[0:36:22] Karthik Ramakrishnan:** And yeah, I think we were working with the ones that, you know, I think are at the forefront of the innovation spectrum in the lab. I mean, there's others, of course, don't get me wrong. It's just we ended up working with the ones we did, but I think the market now understands. Yeah.

**[0:36:43] Alex Bond:** Yeah, yeah, yeah. I mean, it's, yeah, we know the Greenlight team and the Chaucer team really, really well. And so, okay, I couldn't, you'd struggle to find two better places that kind of work with on something, this at the edge of kind of what insurance can offer. But is there, I don't know, we'll have to start to wrap things up, but I was thinking about the way that you could sell. Have you looked at kind of embedded as a potential option in the future?

**[0:37:11] Alex Bond:** Is that something that's being considered?

**[0:37:13] Karthik Ramakrishnan:** I think there's different business, many different business models. Embedded is certainly one, right? like where you can be embedded within an AI vendor's product. And this is a big value proposition for them. So if you look at, I've only spoken of the enterprise, but if you look at it from a vendor standpoint, when that trust deficit exists, how do these, I don't know, what thousand different companies that have been created that are now working with AI and building innovative solutions, how do they get across a trust factor so an enterprise can purchase and adopt their solution?

**[0:37:48] Karthik Ramakrishnan:** And so this is, we provide a path now to say, we will come and assess as an independent third party, and then assure and guarantee that your model will deliver, which allows an enterprise to buy your solution with greater confidence, right, and perhaps faster. And that is essentially an embedded opportunity, a way to look at it. But there's other sort of go-to-market opportunities as well, as I said, from the enterprise. That's a different value proposition as well, where for the enterprise, it's, well, how do I scale this across all of my purchases or adoption of third-party AI models?

**[0:38:30] Karthik Ramakrishnan:** How do we do that in a compliant and confident manner? um like i think the the the we're just at the very start of this journey um on risk transfer solutions for ai some exists um but what needs to exist the future is going to look very different how and this is going back to your very first question where is innovation going to come from uh for for insurance right is it is it just digitization or is it different ways and modalities I think we need to start thinking out of the box how we even ensure these models and what we ensure is going to be another sort of segment of thinking and innovation.

**[0:39:18] Karthik Ramakrishnan:** And how we even capitalize that risk is going to be up for debate as well. I think that there's opportunities to do innovative ways. Like, is there a mutual way of doing this, right? A shared risk pool. Is there alternative sources of capital? And these conversations are happening elsewhere too. But innovation is not just completely brand new. It's how do you even pick up existing things and bring it into a new context and employ that to benefit?

**[0:39:54] Karthik Ramakrishnan:** it, I think all those opportunities exist in this current environment of AI risk.

**[0:40:01] Alex Bond:** Yeah, I agree. And I think what's exciting, and certainly for a person that's worked in the insurance space for 17 years, is speaking to people like yourself. And I think we need to do more of that. And look, I'm not going to bang my talent hat for too long, because I managed to avoid saying that I work in the talent industry for a little while. But that's what we've needed for a long time. We need experts that come in and bring new ideas and then look at what we do in insurance and risk transfer in a different way.

**[0:40:29] Alex Bond:** And I think Armilla is a perfect example of that. I think I could talk to you for hours and we could really get into the kind of the really high level conversation about AI. But thank you for being a guest. I wanted to do this for a really, really long time. So I really appreciate you being a guest on the Leadership Insurance podcast.

**[0:40:48] Karthik Ramakrishnan:** No, absolutely. Thank you so much for having me. I've always enjoyed our interaction. So this is excellent to probably put it on the record. So thank you for having me.

**[0:40:57] Alex Bond:** That's really kind of you. Thank you very much for your time. Appreciate it.

**[0:41:00] Karthik Ramakrishnan:** Thanks, Alex. Bye.
