Go back

OZ Watkin, Co-Founder & CTO at Mastery AI - PIR Ep. 886

from Profiles in Risk

31m 4s

OZ Watkin, Co-Founder & CTO at Mastery AI - PIR Ep. 886

Master AI is a specialized AI platform designed to revolutionize insurance underwriting by enabling real-time portfolio analysis across property, cyber, and D&O lines. It addresses a long-standing industry challenge—the fragmented, non-standardized format of "border of files"—by using proprietary insurance-trained LLMs and agentic harnesses to automatically process, structure, and analyze diverse data sources. These AI agents work collaboratively within defined guardrails to perform tasks like risk aggregation, KYC checks, and sanction screening, significantly reducing manual effort. Rather than replacing underwriters, the platform automates routine, rule-based work, allowing human experts to focus on strategic, judgment-driven decisions. Master AI offers three operational modes—fully agentic, semi-agentic, and manual—tailored to different risk complexities, ensuring flexibility. The company, which joined Lloyd’s Lab after a competitive selection, has validated its product-market fit through partnerships with major insurers and emphasizes the importance of preserving human oversight in critical decisions. By focusing on deep industry specialization and real-time intelligence, Master AI is helping insurers transition from legacy, delayed underwriting to agile, data-driven portfolio management, positioning itself at the forefront of an evolving AI-powered insurance ecosystem.

Transcription

4834 Words, 27113 Characters

English
Hello and welcome to Proposal Risk, this your host Tony Kanye's switch into my radio voice, and today I have with me O's Wacken, co-founder and CTO at Master AI. As always, thank you for joining me today. How's it going? Thank you, Tony, for having me, lovely. Awesome, my pleasure. So we always give the guest the chance to give the elevator pitch. What is Master AI? Cool, so Master AI is an intelligent platform that provides capabilities of AI intelligence into insurance industry. We allow the transition into what we call portfolio underwriting. So in real time or near real time, we allow users underwriters to understand the portfolio exposure that they have on the book and be able to do more efficient underwriting. So we do a lot of the legwork for them using our proprietary technology, which we built over the last 12 months. All right, often when I hear about portfolio underwriting, we're talking a lot about property. That's one of the, or small commercial, but property, it's one of the areas where like managing that catastrophe exposure and what's it called the, I can't remember the exact word, the availability of capital for the risk. So are you guys focused on property or are you guys focused on what type of insurance? We're a line of business agnostic. We're working with a few customers that have 12, 13, 14 lives of business. We support broader than that. So obviously, property comes to mind initially because of the geographical ability to represent catastrophes or kind of exposure to specific areas. But the same thing can be done for cyber, where you would like to understand to which type of industries you're exposed to, which type of professions. The same thing we do for DNO, so delegated of directors and officers. In some cases, actually, we get from customers the combined aspect of property together with DNO, together with cyber. Sometimes they will have different lines of business that ensure the same company, the same building, multiple companies within the same building for all of the three. And what you would like to understand is to shift away from risk exposure, line by line, on one submission at a time, into an ability to really understand and correlate your sort of total exposure to the business. I think that underwriting and portfolio view has not been something new in the industry. We've been aspiring it for the last five, six years. Everybody's been talking about it. The challenge until now was the technology was not there. We were limited by our ability to process, you know, border offides, which come in six, eight months delay. How do we create a more digital, near real time or timely manner of a digital thread throughout the process across the companies? And I think this is where things start to change a little bit, where we're adding, you know, agentic harnesses into the process. We are able to now do agentic underwriting that improves significantly the time in which we're able to generate the border of files, reconcile them against delegated authorities. All of those together, I think, provide now something, which is a little bit different that enables us to have that transition. So let me pause for a second. Okay. So out of the hundreds of references we've recorded, the last three years, all anybody talks about is gene AI. The last year and a half, going on two years, all anybody talks about is agentic. You're the very first person to use the term agentic harnesses. So help me understand that. Yeah. So let me kind of double click on that one. Similar to companies, we don't work as individuals. We work as teams. Teams that are able to cooperate together to achieve a mission in a better way. The successful companies, the combined intelligence of everybody that works in the company, aiming for the same goal. And genetic harnesses in essence are exactly that. Now the ability to take multiple capabilities, agents, pieces of software that have been trained previously on what we're trying to achieve. Just as an example, risk aggregation. We have pre-trained risk aggregation within the insurance industry. You can see the Lloyd sign behind me. We're a bite of the Lloyd lab. But it is really looking into specialty insurance. So if we risk aggregate in specialty insurance, which is quite complex, we are then able to use multiple connectivity between agents, workers, those together get to better results. So if I have a workflow that says, please take something with the innovation. Here is a submission from a broker. The underwriter needs to run through 10 or 15 different steps before they even consider, yes, no, do I want to underwrite this or not. We can run an appetite check, we can run risk aggregation, we can run sanction tests, do a KYC, KYB, know your customer, know your business, do additional analysis. And those together, working together, that's a harness, that's the ability to take together agents and make them work together. Perfect, perfect, perfect. All right. All right, so border roads. This is where we're. The curse of the industry. Yes, this is where being a giant insurance nerd pays off. So I've never actually touched one because I came from primary insurance, not from re-insurance. But I've gone, well, number one, I was educated on them through my educational journey, CPC, UAR, et cetera, et cetera. And then as a recruiter, I have several times recruiter for people who have to mess with them. And I've had people in the podcast who are developing technology for border roads. So yes, border roads. So maybe let's kind of depig the problem a little bit. For those who might not have dealt with border roads. Border is, in essence, a list of your book exposure. All the policies that you have bound, all the customers that you have. What is your level of exposure? What's the premiums that you have undersigned? What's the total exposure that you have for each one of those? So it's a thing about an excelist of all of those items line by line. It gets a bit more complex once you're looking, especially into specialty or large corporates, where there is what we call endorsements. So you have your base policy. And after your base policy, you start endorsing a lot of stuff. You start making changes. So you have different elements of those that each one of them together combines into something a little bit more complex. So border of files, in essence, have this limitation, where they come in Mary shapes and forms, and each one of those shapes and forms needs to be mapped into something that makes sense. Yes, let me post you that. What I've seen from the helping higher data people to build with border roads, and there's no standard. So in commercial insurance, we have the accord form, which is not great, but there's a standard. With border roads, there seems to be no standard, historically. Yes, that's correct. I think there are some attempts in the industry to have regulate some aspects of these multiple times. It has not worked very well until now. In Lloyd's, for example, there is a specific border regulation and standard to how to do that. It is a journey. It's not completely there yet. But what I would say further, that is that once who are insurers or delegated authorities to others to underwrite on their behalf are expected to follow up and understand the exposure that they have on their book based on the delegated authorities. So in simplicity, I'm a re-enture, I'm sending you the ability to delegate on my behalf and underwrite on my behalf. So if I do that, I need to know that you're underwriting what was agreed in our delegated authority. The way for me to know that is based on our border of files. So maybe if we stop for a second the recording, I'm just going to move because there is some cleaning help coming next to us. No problem. Let me. All right. So we pause the recording for a second, move locations within Lloyd's lab, and through the magical video editing and/or time travel we're back here. All right. So we're on border roads and how progress is being made. Yes. And the Lloyd's market does have some level standard. I think it's broader than the Lloyd's market. The Lloyd's market represents a higher level of complexity because of the specialty aspect of it. But it will. it will be in PNS. It will be in ENS. It's in other markets as well. I think wherever you have delegated authority or other people underwrite on your behalf, you will have those border of files. Sometimes even internally within the organization you have multiple companies. Now let's just kind of move away from that as a complexity of a market and kind of unit it into one. You basically get as a company multiple Excel files, multiple CSV files, each one of them is in completely different format. and you need to figure out out of those what other people have underwritten on your behalf. And what is your risk? In the end, in insurance, well, what about risk, right? It is what is the risk exposure that I have? So most companies get it in a leg of a month, three months, six months, and so on, then take another month or two to basically kind of gather all the information into one piece. So at any point of time, their snapshot of their exposure is in a three month, six months delay. Tremendous amount of efforts goes into the process of consolidating those ones in order to better understand your risk. And there has been many attempts in the last few years to try to do that. Now, algorithmically, we can do it one at a time, right? So I give you a set of problems with different type of formats. You are able at some point to figure out how to do that. As you work with many different customers, it is sort of a problem of infinite scale. You are constantly have to adhere to new standards, new structures, new items. So infinitely, you will have to keep codifying something. Because Mastery specifically is using an insurance-based LLM, which we have built proprietary to us, we have been able to educate it on a different type of border of files. So it gives us really strong context of what border is into the model. And more than that, we are able to employ a local LLM that is deployed on a customer private cloud that then studies those CSV files, structures of files. It basically flatten those ones and create the enhanced neural network within it. So it moves the weights based on that understanding. As we do this, it is just about a representation of those ones as the risks are later on. So what we do is we take tremendous amount of data, we train it, then we give it to a customer. A customer gets that model and feeds all of those different CSV formats, file formats into it. And then it asks the model to represent it in a specific way. So it is relatively, it's kind of like you would use HGPT and throw into HGPT in many different documents and then say, hey, build me the following format of information, only specializing in terms. All right, that makes sense. And basically, neural network, all your border of files, and you get this beautiful model, model might not be the right word, but specifically built on your border of files. For your work profiles, for your defined format, correct. Now, if you are using our authentic underwriting capabilities as well, those border of files are generated for you automatically. So the process of our orchestration allows the generation of those border of files for you automatically, while you are transitioning away from the deterministic flows of delegated authorities and others in the industry. Perfect. That makes absolute, absolute sense. All right. All right, all right, all right. Okay, so going back to what I know, we spent a good amount of time on the world. Thank you for further educating me on the border, since it's a piece that I have not touched before, or at least, you know, that have not worked with. I do have a good word it is. Yeah. All right, so we talked about the agentic harnessing and the system being able to do a agentic underwriting. So talking about agentic underwriting, does that mean letting it underwrite the easy stuff? Does that mean we won't need any underwriters? What does that mean? So, at MasterAtlist, we recently actually published a research around the changes in operating models that we believe that is going to happen in the context of agentic underwriting. The underwriters hold a very important significant knowledge within the process in the industry. Our industry is also very much about relationships in the specialty insurance. We believe that underwriters will become more significant in the future, because they will be the ones that sits on the deduction of the decisions. So for us at MasterAtlist, we want to empower the underwriters, but we want to take away the work which is insignificant for them, that is not about them making decisions. Going through a suction test for a customer is not something that the underwriter needs to do. Something that the computer can do it very well. The ability to consolidate information from multiple systems to understand what is the correct format that you need to have in the end for the policy is not something that the underwriter needs to do. We need to bring underwriters to a place where they can focus what we call their cognitive abilities in decision. So they become, in essence, the managers of those agentic capabilities to provide them all the decisions that they need. So what we do is we don't underwrite agentically. We run the process agentically and the underwriters underwrite in the end. We provide them a suggestion for decision together with tremendous amount of context. In many cases, clients actually say, can you guys elevate our more junior underwriters and show them? Beautiful. What is the key capabilities from previously underwritten policies that are similar by our top underwriters? Let me pause here. I love that. Let me pause here for a second. There's two things there. Number one, for the listeners? No, we are not giving them underwriters. We won't have any industry without you. We're making million or multi-million-dollar decisions. We want to human involved in those decisions. Now, we want to give you superpowers so that you can make better decisions. Now, one piece that has been worrying me, and you just addressed it a little bit, and I love that because I have not heard good answers before, is if the better and better AI that we get is used to automate, if we can get to let's say 95-99% pass through underwriting of the easy stuff. How do we train the underwriters to be ready five years from now or ten years from now to underwrite the hard stuff that will always require human judgment? What you mentioned is one way to do that, so I love that. I think it's also we need to observe it in a slightly less obsolete aspect, meaning underwriting simple, underwriting complex. It is about aspects of underwriting within the underwriting process. The machine is to do things that are not simple, but that is more of a machine to do. Collating tremendous amounts of data to understand context is something that we as humans do not necessarily do well, but we do well is on decision in the context of a lot of other things, meaning I know that Bill is from this factory in this location, and he's been underwriting with me for the last 15, 20 years. He knows everybody that lives in that area, so if he brings a deal to me, I know that Bill is good, right? Now the machine should not be deciding if Bill should be underwritten or not. The machine should call it information to understand that this specific property is currently, you know, 500 meters, 500 yards from the beach. Does that represent a specific risk factor within our book that is something that is within our appetite as a whole? And then provide the underwriter to say, look, this is within the context of our appetite, or it is not, this is within the boundaries of what we want to underwrite as a global company for the book that we have. In the end, it is the decision of the underwriter to say yes, no, and at what price? So if we take a more junior underwriters and we say, hey, our top underwriters in the last 10 underwritten policies have confirmed the following price or the following sanctions or the following endorsements on specific policy or these conditions learn from those. Now it doesn't mean that you should copy paste them. It means you have a consideration that those more senior underwriters have done that. Yeah, it was very similar. Sorry for interrupting, but it's very similar to me as a middle market underwriter with Liberty, which was my last underwriting role, where I've got my CPCU, my bunch of designations, and a year of small commercially underwriting experience. And now I'm underwriting middle market, but I'm sitting next to skip that is literally the guy's name, who has been doing this for 50 years, literally retired from Liberty after 50 years, and being able to run something by him. And yeah, I'm not going to do exactly what he tells me to do, but just yeah, being able to watch the game video, right? Being able to see what he's done in the past with risks like this will like it is. And remember that skip doesn't always have time for the junior underwriters because he is underwrite as well. He has his own PGL, his own underwriting target. But if we can take skip knowledge. compound it into an ability to represent a reflection to those underwriters and then we can elevate them. As you said, skip is another data point. It might be an important one or less important one depending on the underwriting specific journey that you have had or the specific submission that you are dealing with but it's another data point is important so it's the combined intelligence of the organization that you can now leverage on that. I think in general, they're going to be a consideration of what is this new operating models in the context of agentic capabilities? Yeah, so I think first that agentic is quite a buzzword so I'll move away from that for a second. When I say agentic, we refer at least at mastery to three things within the team. One is we are providing a piece of software capability to have agency to achieve a specific goal. For example, goal risk aggregates the following information. So I give it an agency and it goes on a specific goal that I'm trying to achieve. Two is it works within guardrails. We are regulated in the stream yet. This is not a wild west. This is not me operating in a sort of personal manner of goal. I aggregate my Gmail and add some information and send whatever you want. Everything has to be working within guardrails. The same like you as an underwriter before that. You had to go and work within the boundaries of how you're supposed to work. The third item for us, so first one was agency to achieve a goal. Second one was walk within guardrails. The third one is learnable. We have developed specific recursive learning capabilities which looks at the information that is running through the platform, learns from those ones and then makes suggestions to the underwriting or chief underwriting team to go and reinforce the capabilities that they decided they want to reinforce. So it's governed reinforced learning capabilities. These are for us, when I say "agentic" I mean these three capabilities in a box. I believe that the industry as a whole will have to look into what are the new operating models. Where do we employ judgment of people and we want the human in the middle to make decisions and where do we want for things to be done and run automatically. So in the case of mastery we have developed within our capabilities three options. Option number one is I would like to run this specific step completely agentically go, run it, finish what you need and then come with the information to the next step. So if you fulfill specific goals. Option number two is run what we call semi-agentic. So come complete the step at weight. Someone will go and check what they want to do and then they will go to the next step. Very typical is pricing for example. Conclude all the pricing models, run through the information, show me the price, show me the information and I will make a decision of how I want to pursue. The third option and that's typically on let's call it the negotiation steps post pricing that is completely manual. So come to that step, hold. I will do what I need to do, I will negotiate with skip, I will negotiate with the rest of the brokers. Once I'm done I'll confirm it and I can go into the binding process or into the co-operations. So there is these three different options within this new operating model to go and execute based on those ones and I think that this needs to be explored. There are subjective to an organization of how organization works, the teams that are there and how to best reorganize around those capabilities. In the end the industry especially in a soft market is challenged with the cost and efficiencies of how to execute things and further than that if we want to scale and grow we have to bring the unit costs down. So especially in new products that might be more competitive, might be more repeated, slightly less specialty. We need to be able to keep costs very much down. I think also consumer behavior is going to change over the next few years so people will start purchasing their property insurance through JGPT, through Claude. They will type oh I just bought a property, what does a property insurance look like in Savannah and how do I now ensure it? That will propose new models so on our side we want to prepare systems to be more agentic friendly and to be able to communicate between different agents on the purchasing side on the channel versus also the agentic submission or underwriting process. I am always curious on the origin story. Look at your LinkedIn, it does not look like you grow in insurance. So I am very curious how you realize that this made sense for insurance and the world great people to build it. It is interesting and on the spot I think so I have gotten into insurance in the last few years. I started in kind of building tech and then moving in between different companies, mostly around telecom and financial services, previous life. I believe the problems are quite similar in the end, those are quite sort of industries that have been developing over many many years. Both of the financial or FSI services and in telecom that has been around for the last 100 years or so, slightly more young than insurance. But in the last few years through different consultancies and consulting companies have worked a lot on insurance implementation and that is really what led us from a story perspective to mastery. We have been doing a couple of transformations about a year and a half ago, two years and we have seen that those digital transformations just don't work. They don't take it up. It takes you a year or so to kind of identify what an organization needs, another year or two to run your RFPs and selection of vendors, another couple of years to then implement a large system that does a complete workbench and all these other stuff. And in the end of it, the KPIs and ROI becomes more of an art than science. So as a large company, you've invested 100 million on changing a workbench and in the end after those five or six years, it's no longer relevant because the industry has moved away and all the capabilities are pretty redundant to be honest. So I think there has been an aspiration from mastery side to develop what we call self-healing intelligence layer and that's kind of the aspirations with the product and where we are heading in building it. All right. Very interesting stuff. I like it and if I'm not mistaken, based on on LinkedIn, you opened this particular company six months ago and you're already on Lloyd's Labs. So yes, we started really with thinking about the idea about 12 months ago and then kind of went into initial interview stage. We've spoken with a lot of sea level executives, really trying to understand where they see the technology heading, both from a business and technology side. We tried understanding the constraints of how because you spoke about generative in the beginning of our conversation. I think generative has been around us for a while now, a couple of years, probably three years since the sort of GPT moments. There is more adoption within enterprise for generative, but generative by itself has little premise to kind of fulfill what AI really could do in organizations. So we've spent quite a lot of time really thinking about not to best engineer this. We've collected the team of exceptional, exceptional ninjas that are far smarter than me, to be honest, that allow us to build this really sophisticated technology. We have focused a lot on the insurance industry, building our own insurance models, building our own orchestration, our own agentech harness within that. So highly specialized for insurance. We believe in specializing in the end. That's how you bring true value and that generalistic sort of models that runs all across. And yeah, we spoke with Lloyd's lab team about three or four months ago, just a little bit after we started, applied. There has been about 500 companies that applied. They took the top 15 for a pitch day, very similar to sort of investment, traditional investment stuff. We pitched and got selected in the top 10. So now we're part of the Lloyd's lab. I think that the Lloyd's lab is to clarify to those that do not know it very well. It's a great opportunity from a really large 60 billion market within the insurance industry of specialty and complex insurance to bring companies that are really already with a product that is in its early phases of the product market fit. And trying to, it's voted by the market executives as well. So the market executives decide I would like to work with that company. And based on your sponsors, you are being selected or not. So they want you come and you start working with six or seven or eight different companies. So we've moved very quickly into working with really large companies. They want which is a great endorsement for us both on the belief of the product that we have built, but also making sure that the product market fit is there. Together with other clients that we have accelerated both in the US and in Europe, it's been quite a journey and we feel very lucky in the way that it's been going so. Fantastic. Bose, thank you for your time today. This has been a super interesting look forward to having you back on the show. You just let me know. Normally what I say is, I'd love to have you back in a year or two. You guys are so early on and I'm sure changing so quickly because of the loads of opportunity. Yeah, just let me know when you're ready to come back. I'd love to have you back as you started deploying this. Yeah, thank you so much for your time today. We pleasure really appreciate it and yeah, the market is changing really quickly so we're part of this wave and feeling humbled to be part of it so.

Podcast Summary

Key Points:

  1. Master AI is an AI-powered platform that enables real-time portfolio underwriting across multiple insurance lines, including property, cyber, and D&O, by unifying fragmented data into a coherent risk view.
  2. The platform uses proprietary insurance-specific LLMs and agentic harnesses—teams of AI agents that work together to automate complex workflows like risk aggregation, KYC, and sanction checks—while operating within strict regulatory guardrails.
  3. A core challenge addressed is the lack of standardization in "border of files," which traditionally require manual, time-consuming consolidation from diverse, unstructured formats like Excel and CSV.
  4. Master AI trains custom models on each client’s historical data to flatten and understand unique file structures, creating a tailored, context-aware intelligence system that evolves with new data.
  5. Instead of replacing underwriters, the platform empowers them by automating routine, rule-based tasks, allowing human underwriters to focus on high-value decision-making and judgment.
  6. The system supports three operational modes
  7. The company was founded after recognizing that traditional digital transformation in insurance fails due to slow implementation and outdated technology, leading to a focus on specialized, self-healing AI for real-time decision support.
  8. Master AI is now part of Lloyd’s Lab, validating its product-market fit and accelerating adoption with large, complex insurers in both North America and Europe.

Summary:

Master AI is a specialized AI platform designed to revolutionize insurance underwriting by enabling real-time portfolio analysis across property, cyber, and D&O lines. It addresses a long-standing industry challenge—the fragmented, non-standardized format of "border of files"—by using proprietary insurance-trained LLMs and agentic harnesses to automatically process, structure, and analyze diverse data sources. These AI agents work collaboratively within defined guardrails to perform tasks like risk aggregation, KYC checks, and sanction screening, significantly reducing manual effort.

Rather than replacing underwriters, the platform automates routine, rule-based work, allowing human experts to focus on strategic, judgment-driven decisions. Master AI offers three operational modes—fully agentic, semi-agentic, and manual—tailored to different risk complexities, ensuring flexibility. The company, which joined Lloyd’s Lab after a competitive selection, has validated its product-market fit through partnerships with major insurers and emphasizes the importance of preserving human oversight in critical decisions.

By focusing on deep industry specialization and real-time intelligence, Master AI is helping insurers transition from legacy, delayed underwriting to agile, data-driven portfolio management, positioning itself at the forefront of an evolving AI-powered insurance ecosystem.

FAQs

Master AI is an intelligent platform that enables real-time portfolio underwriting by automating complex data processing and risk analysis. It helps insurers understand their total exposure across multiple lines of business, overcoming delays and inconsistencies in traditional border of files.

Master AI is line-of-business agnostic. It supports property, cyber, and D&O (Directors and Officers) insurance, allowing companies to analyze combined exposures across different insurance types within the same building or organization.

Agentic harnessing refers to combining multiple AI agents that work together to achieve a specific underwriting goal. These agents perform tasks like risk aggregation, KYC checks, and sanction screening, enabling faster, more accurate, and comprehensive underwriting decisions.

Master AI uses a proprietary insurance-based LLM trained on diverse CSV and Excel formats. It learns to flatten and interpret these files, creating a customized model for each client that accurately structures and analyzes their data in real time.

No, AI does not replace underwriters. Instead, it automates repetitive tasks like data consolidation and risk aggregation, allowing underwriters to focus on high-level, judgment-based decisions such as pricing and policy approvals.

Master AI learns from past decisions of senior underwriters and provides context and suggestions to junior underwriters, helping them make better decisions by understanding patterns and best practices in similar cases.

Chat with AI

Loading...

Pro features

Go deeper with this episode

Unlock creator-grade tools that turn any transcript into show notes and subtitle files.