AIP Podcast EP 66 - Deploy AI Confidently With AI Insurance & Risk Management by Armilla AI

This episode's guest, Karthik Ramakrishnan, CEO of Armilla AI, shares how he has worked globally to address the risks involved in the development, deployment and implementation of AI and launched Armilla AI, an insurance solution that directly addresses the risks of AI, to help companies protect their interests in the fast-approaching age of AI. Armilla AI is a provider of AI risk mitigation and transfer solutions on a mission to enable enterprises to safely deploy cutting-edge AI. Using industry-leading AI/LLM evaluation technology, the company assesses and measures the risk level of AI models to provide AI verification and warranties.

Full Video: https://youtu.be/U1kOFhJLH9k

Learn more by checking out the links below.

Follow AIP Affiliate, Armilla AI
Website: https://www.armilla.ai/
LinkedIn: https://www.linkedin.com/company/armilla-ai/

Follow AI Partnerships
Website: https://www.aipartnershipscorp.com/
LinkedIn: https://www.linkedin.com/company/aipartnershipscorp/
X: https://twitter.com/AIPartnerships

The AIP Podcast is hosted by Anne Cheng, on behalf of the AI Partnerships, a Railtown company

Transcript

SPEAKER_00 0:05

The risks involved in the development, deployment, and implementation of AI have never been greater. From high-profile leaks of internal data by large corporations to entire courtroom battles with false evidence created on deep fate. But the solutions to ensure these risks are incredibly few. Today we have a guest, Karthik Ramakrishnan of Armila AI. He has worked globally to address these risks and launched Armila AI, an insurance solution that literally directly addresses the risks of AI to help companies protect their interests in the fast-approaching age of AI. Well, Karthik, it's so good to have you on the show today. Tell us a little bit about you. How did you go from tech to actually addressing the risks of technology?

SPEAKER_01 1:02

Thanks for having me, Ann. Great to speak with you. And I think what we're doing today is an evolution of everything that's transpired in the last couple of decades of my career. So I got into AI about 10 years ago, or even before that. Being in Toronto with the breakthroughs in artificial intelligence, starting at the University of Toronto with Dr. Hinton's lab and everything back in 2012, it was very hard to avoid that excitement that existed. So I got into the space of AI. This is my third AI startup. Started doing it before it was, I guess, cool. And then most recently, before Armela, I had the opportunity to work with Dr. Joshua Banggio. And there we started building some of the first enterprise applications with AI at Element AI. And we were working with very, very large kind of top companies like Toyota and HSBC. And while we were talking about why enterprises should adopt AI and how it would transform them as businesses in the coming decades and how they had to start in that journey, one of the fundamental issues, and it continues to be so today, it was certainly back then as well, which was AI is unreliable because it's a probabilistic system. And so it's not a matter of if but when an AI system is going to make a mistake, right? It approximates towards an answer, unlike deterministic software, traditional software, which is you code it to do X, it will do X, versus an AI system learns what the output should be and it approximates towards that answer, which means there's an uncertainty to that answer, every answer. And no AI system is going to be 100% accurate. But if you're an enterprise and you're going to hand over decision-making authority to such a system, enterprises need to be very comfortable knowing when a model is going to do well, but more importantly, when it's not going to do well, when that approximation of that answer falls below a certain acceptable threshold of accuracy. And to do that, one had to understand AI at a deep level. How does it work? And how do these systems get built? How do these models work? And that was our expertise. And we said we are best positioned to actually answer that question and solve for that problem. And so Armilla became sort of the vehicle for us to answer that question. How do we help enterprises adopt AI in a confident manner by increasing the trust in these systems, by helping them understand where these systems will fail and which translates into risks? And that's how I got into the world of risk and risk management around AI and now ensuring those risks as well.

SPEAKER_00 4:20

That's amazing. That's an impressive background. Tell me a little bit more about Ermila and how it works, and specifically, how does it underwrite the risks of AI? And what are these risks that you're attempting to address?

SPEAKER_01 4:35

That's a great question. So very much like you know, um cyber insurance. Maybe I'll tell a little bit of a history of insurance in some in some ways, right? It is the story of how insurance got started. Um, you can go back thousands of years into back into the Hammurabi code uh in Mesopotamia, when you have the first insurance contract that was written on a tablet or etched onto a tablet, which basically said, you know, the the king would uh would underwrite or help cover the risks of trade that happened over the seas. It'd start there. Or you can go back 150 years ago when electrification uh of uh buildings and electrification of the economy started, and fires used to be a big concern, right? Um when these electric systems grids, which were pretty nascent and very rudimentary at that stage, when they would fail and you have fires started. So you started something called the underwriting laboratories. Underwriters came in to understand the risk of these electric systems and grids and help ensure uh those properties, right? And now that enabled, in both those situations, trade to flourish because you had someone taking on the risk. You had electrification to flourish because you could underwrite that, you can turn that to auto. Cars came into the scene uh you know back 140 years ago, and now we have um automobile insurance as a way to underwrite those risks of what automobile accidents could entail. And so now we have the next stage of technology which is coming, which is artificial intelligence. To enable artificial intelligence to flourish and get adopted well, you need to understand these risks. And so we started with that premise. The best way to understand those risks is to be able to technically understand how these risks manifest themselves. That's the philosophy. And the way we do it is both in a quantitative and qualitative uh set of factors. So we actually go and test these systems across a varied set of dimensions, right? Starting from the use case, the domain it's implemented in, the regulatory environment in which it operates, um, the governance procedures that a company has in developing, deploying, and maintaining these models. And then finally, at a qualitative level, we then go and test these systems with those factors to say, is it accurate? How robust is it? How resilient is it? We look at the data, we look at the security protocols, we look at the biases, if bias is a concern in that use case. And all of this then gives us a very critical quantitative level understanding of that risk of a model, right? And it's very comprehensive. In fact, if anyone in the enterprise has bought cyber insurance, you would know that cybersecurity is one of the key underpinnings of being able to underwrite the risks of cyber threats and risks, right? So we're taking a very similar approach and we're creating some of the first standards in the insurance industry for how to underwrite, understand, and then underwrite these risks of AI at a quantitative level.

SPEAKER_00 7:45

Got it. It's really quite revolutionary, isn't it? And yeah, it really does make sense. How did you come to this aha moment and what other types of risks are you seeing developing with the new forms of AI being launched on the horizon? Like things like multi-agent frameworks or social intelligence robots for psychotherapy?

SPEAKER_01 8:07

Um, it could be a very long answer because when you look at the dimensionality of AI, there's so many areas in which risk can come in, right? Um you have uh at a at a regulatory level, right? You have it's not a regulatory risk, but how do you comply with the regulations? And these regulations are changing. But I guess more importantly, the technology itself is changing right before our eyes. Every three to four weeks, there's a breakthrough that happens. Uh, you know, the deep seek revolution. Prior to that, you had the Chat GPT revolution, prior to that you had the neural network revolution. Each one of these creates a um a risk in that we don't understand this technology, but as enterprises, they have to keep up, they have to adopt. Now it's not a choice whether to do AI or not. You have to rethink how your enterprise functions and you have to adopt this. So that's the first premise. You have to adopt this technology. Therefore, you need to understand um the risks inherent in the model itself. You have to understand the risk at a product level, right? How will this product fail? Uh, there's technical factors of failure, but there's also um uh design level factors of failures, right? And now that brings the question, as you said, right? You had simple models before, single model solutions, then you had an ensemble model of solutions, now you have autonomous, completely autonomous solutions, the agentic level things that we're talking about. And multi-agents means these agents will interact with each other in an autonomously. So you have degrees of complexity that are being put in, and so degrees of risks that are increasing. And so uh there's technological risk, there's business level risk, there's regulatory risks, um, and and we really have to also think about societal risks in some ways as well, right? So, how do the solutions that we put out affect our stakeholders? When the model makes a mistake, it's not just a mistake that it made. It could cost billions in dollars of revenue, it could have a monetary or financial impact to your stakeholders, it could have um um availability of services, uh risk, right? So you you know, you could go to the extreme, right, to decide a level of you know, um um image recognition models that could misidentify someone, or you could have credit lending models that may deny claims to uh uh loans to someone, or uh a GPT system that frankly just provides bad advice. And we've had the examples of that just two days ago, where cursors, chatbot, cursor is uh if you're not familiar, cursor is a uh platform that allows uh easy code to be written. The platform generates code that developers can pick and choose and run with, and then coding time gets reduced. But that's cursor. The chatbot basically advised saying to some developer that you know uh the policy level of uh cursor had changed and therefore they've been denied services, or they couldn't do X because of that. It was completely made up, right? And so now you had a bunch of developers who are like, if that's the case, we cannot work with cursor anymore. Think of the monetary impact to the um to cursor themselves, but the um functional impact to the users of cursor, right? So when these failures can happen, you really need to know not just what these risks are, but also how are you gonna put guardrails around it? There's design level risks or guardrails that you need to put in to the design level, you know, how are you catching these errors? Have you caught them all? And AI being probabilistic, it's gonna say answers that you've never seen before. And so now at inference time, how are you catching these errors as well? Right? So these all come down to good design of AI systems. Um, and so now you have design level risks as well. So um this is sort of when you as an enterprise or one as an enterprise is looking to implement AI, uh you're you're you're not only dealing with the changing nature of the technology itself, but the changing nature of the products, the changing nature of applications and use cases, and then designing your your risk parameters or identification of risk parameters, and then controls around these risks as well. Now, the final leg of all of this is you're never going to be 100% risk-free. Right? No matter how much you do, you have a residual layer of risk. And that's where insurance steps in to say, look, you've done everything right, possibly within your control parameters. Now let's also protect you for the unknown unknowns. So risk comes in three different ways. You have risk identification, risk mitigation, and then finally risk transfer, risk protection. And those are the three things that Armala is trying to do.

SPEAKER_00 13:00

That's amazing. Um, one further out, I know we're kind of out of time, but in this world where AI is being adopted, left, right, and center, what kinds of organizations do you think are most at risk from the progression of AI applications and why?

SPEAKER_01 13:18

Um I think every organization, regardless, and I can it's an easy answer if I said regulated industries are highly at risk. Why are they regulated? Because they have a high level of consumer impact, right? So healthcare, financial services, uh maybe even manufacturing to a certain degree, uh highly risky or high-risk uh environments. But if you look at something called, you know, risk managers look at something called materiality. So you could be in retail, you could say it's I just have a simple marketing chatbot, but that has a high touch point with their consumers that if it says something wrong, there's a brand risk, there's a reputational risk, and of course, uh legal risks associated with bad outputs as well. So materiality of the use case matters. So materiality is a concept that you use to identify the risk exposure or risk level. Uh and if you look at a use case standpoint, combine that with your uh type of company, the size of company, a very large company has a high risk exposure versus a small startup. Again, your risk exposure is smaller because your revenues are smaller. Don't have that many clients, the impact is lower. So when when you take it from that lens, um I'm loaded to say one industry is more at risk than someone else. I think it comes down to the use case and then an understanding of the impact that your use case has on your stakeholders.

SPEAKER_00 14:40

That's been that's very useful for all of us here. Well, Karthik, I think the world can truly learn a lot from you and your story. I've truly had a blast on our show today. And once again, to all our listeners, my guest is Karthik Ramakrishnan from Ermila AI, and I'm Anne on behalf of the AI Partnerships Corps for the AIP podcast. That's all the time we've got today. If you've enjoyed our show, please do follow, share, and subscribe to our podcast. You know that it does so much more to educate our audiences about developments in the AI world. Take care, everyone. Over and out. Thanks for having me, Ian.