Speaker #0FELIX: Welcome to the Akur8 Academy Podcast, brought to you by Akur8. Akur8 Academy is your free actuarial training community for life and non-life insurance professionals globally. This is where actuaries, insurance professionals, and industry leaders explore the ideas shaping the future of the profession. Every insurer is being told AI adoption is inevitable, but most of that adoption is still stuck at the pilot stage. In this episode, we go inside how Akur8 rebuilt its own processes and culture to move past that stage, the mistakes that cost us productivity along the way, and the changes that actually got AI working at scale. My name is Felix. I'm chief operating officer at Akur8, and I'm joined by Ludovico, who is Akur8's head of AI transformation. So for those not familiar with Akur8, Akur8 is the leading AI actuarial platform for both P&C and life insurers globally. And Akur8 Academy is our platform dedicated to actuarial learning. When Ludovico and I sat down to prepare for this podcast, the question we were really struggling with was what original content can we bring to an insurance audience that will actually help you think through the challenges you're facing with AI? And we all know that AI is being adopted in insurance. Everyone has now accepted that it's coming. Almost everyone intends to adopt the technology, but in insurance, adoption is still in its early stages regarding actual in-production use cases. And meanwhile, we're all being flooded with the same content that's not always very actionable. So what could we say that's meaningful and different? Well, Ludovico and I have both been very personally involved in Akur8's own internal transformation as a software company into an AI-driven organization. And so we thought, well, why don't we just tell Akur8's personal story? Because Akur8 is actually part of two industries, the software industry and the insurance industry. And as a software company, we've had to go through a significant transformation in order to fully adopt AI. So, what we're gonna do today, that hopefully is a little refreshing compared to all the other AI content out there is tell you the personal and honest story of how Akur8 transformed internally to adopt AI. And from that try to extract some learnings that are relevant for insurance companies and for actuarial departments more specifically. So to tell Akur8's story, it's worth starting with who we are as a company and our DNA. Akur8 has a strong culture of innovation. The company was founded on the premise of bringing the latest technology to actuaries. And Akur8 successfully pioneered a lot of new technology for insurance, like machine learning applications for insurance pricing specifically. And we've built over time strong expertise in the fields of machine learning and cloud engineering. And I'm not saying this to sell Akur8 to you. I'm saying this because even as a company with these foundations, AI adoption for us was hard. It didn't come naturally. It required a lot of trial and error and a lot of preparation. And we started our journey, like most companies, a few years ago by providing enterprise access to an LLM provider for all of our employees. And we encouraged adoption and we encouraged experimentation. But after a few months, the adoption and use stayed relatively superficial, and we weren't seeing a deep impact on the productivity of the company. And we soon realized that the culprit for that was our own internal processes. And even as a relatively young seven-year-old company, we already have deeply established processes for building and for shipping our products. And there were many cases where these processes were in conflict with the new technology. LUDOVICO: Yeah, Felix, I can add an example on that. Very early in this transformation, we have one of our top performer engineers, a guy who really likes to try out new things. And you know, this guy made a very important feature for us, a feature that a lot of our clients are requesting, a feature that allows to view and manipulate data more easily on our platform. This person designed and built this feature in just five days. And this was very impressive. However, like it was working, but we could not get that to our clients because software is not just building something as fast as you can. It's also making sure that software works. And we have a lot of processes that are built and ensuring the quality of our software. And we found that actually this feature was not compliant with all the checks and validations that we have. And even internally, our team did not trust that this feature was up to the standards that we have. Very good individual initiatives were a bit blocked by these organizational processes and bottlenecks. So processes that we set up before AI existed, which served us very well for a long time and which we're very proud of, but which are not adapted now to the new world. This is really what spurred us to say, what can we change organizationally so that we really leverage AI and not only AI, but our most productive employees. FELIX: Yeah, Ludo, I believe that this example shows that AI is not magic. The surrounding process still matters. After, let's say, about a year of experimenting with AI, we realized that there were three main things that were necessary to fully adopt the technology. The first is regarding the process. So what we just discussed. If you don't rethink and re-engineer the process that you're trying to improve with AI, it's not only that it's not going to work correctly and you won't see the associated productivity result, but it might actually harm overall adoption within the company because if the process isn't rebuilt, then people will think the AI is dumb, that it doesn't work, and that will harm trust and adoption globally. Two is organizational change. So we had to create new roles, new entire teams and also adjust a lot of existing roles. And the first example of that is that we created a whole new AI transformation team from scratch. Ludovico is going to go deeper on that in just a minute. But we also had to redefine or adjust a lot of existing roles. And we estimate that about thirty to forty percent of all roles at Akur8 had to evolve over time. A good example is designers and engineers. With AI, designers now have the tools to actually go into the code and make changes, something that they didn't used to do or be able to do. And on the other hand, with AI, engineers spend much less time on coding. And so they have more time to actually think about features. And the organization needs to adapt to allow people to cross those boundaries. And so we extended the roles of these people to include these new tasks and activities. And number three is regarding change management. We started with a very bottom-up, non-mandatory, employee-driven type of adoption. And that's good because you need the teams to feel like they have the freedom to experiment. And that curiosity needs to be encouraged because in the end, these new ways of working with AI are going to be figured out by the team, right? Not by management. So that needs to be preserved and encouraged. But we quickly realized that wasn't enough because you inevitably have some people who will resist the change. It's human, it's natural. So that will slow adoption for the rest of the company, even for the early adopters and ones that are all into AI. That's why we had to complement this bottom-up approach with a mandatory guideline to adopt AI and a central team that was built to guide, to enable and to co-build with all the different teams. And Ludovico is now gonna tell you more about this team that he's now leading. LUDOVICO: Thank you, Felix. So we built what is the AI transformation team. And the purpose of this team is really to guide the change of the organization throughout all of our functions. So the team was assembled by a lot of different people. They come from the engineering side, but they also come from our business performance team, from our sales operation team. So it's really to drive the initiative across the whole company. They're really like our top performers, our best people across all the functions. So as Felix already said, we are trying to combine a top-down and a bottom-up approach. Because you know, people just don't know AI, right? AI is a new thing. Essentially, you know, nobody knows really like where we're going to end up in AI, right? A lot of people are divided between maybe the people who think that AI is just, you know, some stock market bubble and people who think that we're going to have artificial general intelligence in a few years and everything in between. So of course it's difficult for any individual to navigate all of this. And the AI team is really the one who is doing the investigations, who's trying to understand what is the reality on the ground, a reality that changes. AI evolves, it is roughly doubling in power every seven months. So rapidly shifting landscape. And we're here really to help all the teams orient themselves in this landscape and provide all the latest up-to-date tools that can be provided through AI. And bottom up is because even if we are the experts of a lot of functions, we're not always in the fray. We notice that the top performing employees, the top performing engineers or sales execs, they have their own methodology, right? You know, some intangible quality. And sometimes that intangible quality that makes them great can be replicated through AI. You can see what their process is and try to rebuild that through AI agents. And of course, we're never going to replace our top-level employees. We want our mid-level employees to be more like them, right? And so what we do is we take the best that we can find in our organization and find ways to serve that to the rest of the organization. Now, all of this is an important technical challenge. If you're doing software engineering with AI and if you're doing actuarial science with AI, there is a technical challenge in understanding how to plug in all the various subsystems together. But we realized very early that for us, this was not at all a technical challenge. It was going to be mainly a cultural and people challenge. We want the organization to evolve maturely as a whole. And that means overcoming all the skepticism, maybe a lack of actual skill sets for people to use AI. And we identified this was really, really the most important challenge. What we saw is people telling us things like, I don't trust AI, right? A lot of people are just skeptical that AI can produce good code. In the early days, perhaps three years ago, we were already told by maybe OpenAI or by Anthropic that AI was going to replace software engineers. Three years ago, a lot of people tried out AI and saw that it was not up to scratch. And of course, once you create an opinion, it's difficult to change it. So it is really about having an honest conversation about what are the trade-offs with using AI. There is a trade-off as many things in life, right? You are able to go faster by using AI for sure. You lose some understanding of what you are doing. Instead of doing it yourself, you delegated it and now you don't know what has been done. Another objection is that people, you know, they just didn't have time maybe to explore AI. Often our engineers, especially our sales execs, have a lot of pressure to perform, a lot of tight deadlines, taking a risk on something that you have never tried before, you know, offers a risk to yourself, a risk to your reputation, a risk to your compensation package if you are a sales exec and you do not manage to use AI the way you wanted it to. And so again, what our team did was build a lot of enablement programs, a lot of small team coaching sessions. We mandated a lot of time across the company for AI exploration and building to drive the adoption at scale, because in the end the organization evolves if everybody improves 10 or 20%, not if only a few people do. Now let's go a bit into how this looks technically on the software development side of things. So as we know, AI output by itself is really powerful, but at scale it becomes unmanageable. And this is due, one, to technical limitations. I mean, AI has limited context windows. After a while, if it tries to ingest too much data, it degrades its performance. And again, if you just leave AI running long term, eventually it will become such an interconnected mess that you will not really be able to understand it. We really cannot afford to just lose control of our software to not understand it anymore. It was never like a possibility for us just to say the AI is going to write all the software, even if it was technically feasible, because we cannot afford to lose control over what for us is really business critical. And the term that is evolving, you know, AI augmented engineering is actually constraint engineering. And constraint means you as the human, as the engineer, as the domain knowledge expert, can put a lot of boundaries. And split your complicated system into a lot of subsystems within which maybe AI develops everything. And so your subsystem becomes a bit of a black box. But as long as you understand the interconnectedness between all the various functions, you are able to map this. And the way you box AI is having human engineers really designing these boxes, these interfaces very cleanly. And then add a lot of software deterministic checks. So those are like really test cases, traditional software engineering cases that run a lot of deterministic numerical quantitative checks on all these boxes. And as long as those checks pass and those checks are reviewed by humans, we are then sure that the subsystem is acting as we expect it to act. And so it's called constraint engineering because it's really about like thinking of like what constraints do I give the AI, right? I don't let the AI just evolve on its own in any possible direction. I really give it strong checks that I control and that I know well. And this way I manage to understand how a system works. This is now how we evolved our full software engineering process. What we have is a set of AI agents. It starts from taking the customer feedback, evaluating it against the corpus of our knowledge about the insurance industry and about our software. An AI agent builds a possible solution to the customer's feedback, to the customer's problem, which our product manager, our product team reviews and then uses AI to design a solution for. Once that solution is designed, it goes through a set of steps, each managed by an agent, but crucially with engineers at each point, approving, managing the split, and reviewing. We built all this process, and even today we already know that compared to our pre-AI baseline, we are already seeing a 50% increase in productivity. So this is a huge gain for us, of course. And yes, we're very proud of this. FELIX: Thanks, Ludo. So you can see that even as a software company with a culture of innovation, the road to adopt AI was not straightforward. And I would argue that for insurance companies and in particular actuarial departments, they're all facing all the challenges that Akur8 has faced and that we've mentioned, but are also facing an extra set of challenges given the particular nature of the insurance industry and of actuarial work in particular. And there's three main reasons we could think of at least why actuarial work is fundamentally different and specific. One, errors are spotted much later. Of course, if there's a bug in your software, it gets spotted maybe after a few days, it gets fixed after a few days after that. But if a carrier makes a mistake in its pricing algorithm, it can be years before the consequences actually materialize. And it takes much more time and effort to fix the problem. Number two is that while, you know, a language model performs extremely well to write code, these models can't really, I mean, at least not yet, they can't really reproduce actuarial judgment or a deterministic model that has been adjusted and refined with actuarial judgment. That's not yet the case. And three, carriers of course are highly regulated, much more than software. And that constrains the actions they can take and really adds an extra layer of complexity when adopting AI. But I think that despite these differences, the goal we're pursuing is the same. Actuaries want AI to help them become more efficient, do more and do better work. And chief actuaries want the same thing. And the key question is how do they make sure AI stays accurate, stays on target, and can be verified? And we think there are two main answers to this. One is good infrastructure, and two is good guardrails. We're going to talk about these two points, infrastructure and guardrails. And let's start with infrastructure. What do we mean by infrastructure? We mean a unified data foundation underlying the AI models. We mean the business logic and context that it will have access to, and of course the security component of keeping AI safe. And the point on infrastructure is really one that's very linked to a question we're all asking ourselves right now. Should I build or should I buy the AI technology I want for my team? And is it mission critical for me to build my own infrastructure layer? It's a very timely question because of course building our own tools and applications has, you know, indeed become much easier with AI. However, building the infrastructure and ensuring it's scalable, safe, and accurate that has not become easier because that type of work still can't be easily outsourced to a language model. And so effectively, AI has shifted the burden and the bottleneck from application development to infrastructure and security. We won't pretend that we're unbiased in the build versus buy argument. Of course, we're a software vendor. And in any case, the answer will always depend on your context, on your own goals. But it is our ambition at Akur8 to serve as the infrastructure layer for actuaries. We're constantly investing in applying the latest frontier models to actuarial workflows so that actuaries don't have to do that work. We give our customers enterprise grade security, reliability, and transparency. And on top of the infrastructure layer, AI allows Akur8, as Ludo explained, to develop features much faster. And those productivity gains are all passed on to users in the form of more powerful features that are delivered much quicker than before. So that's on the infrastructure part. The second really key component we think is crucial for actuarial departments is the guardrails that you set around your AI. LUDOVICO: Akur8 encodes all the deterministic models, the model buildings, the KPI creation, machine learning algorithms, has a governance layer, so everything is traceable, has a security layer, so everything is secure. But these deterministic models allow also the AI to use and to check its own understanding on something deterministic. The basic example is you know, you cannot ask a large language model to calculate an average of a large data set, right? It's not going to just read that and figure it out. What it will do is it will run a Python script or some other execution layer which has the mathematical very simple algorithm of calculating an average, return that to the AI model, and the AI model will then say, the average is this. So, of course, if you bring that to much more complex machine learning models, you will need all this deterministic execution layer. But not only that, your context, your data, your institutional memory, your company strategy is also a non-deterministic, but still a constraint, a check that you can give. Because what you can do is you can have agents now being adversarial towards one another. So you can have one AI agent doing the work and the second AI agent using your business strategy, your data, your institutional memory. So I think this is like a crucial thing that we'll see all the industry evolving in because for the first time you can in some way through LLMs codify more complex, a bit more intangible business logic. Business logic that before was only verbal, it could only be verified by a human, but now it can be described to an AI, and the AI can do this verification. So you can encode this sort of business strategy, institutional memory. Then of course, the whole execution layer is important because you want everything to be auditable. You want to be able to go back in time and see you know, three years ago, this action was taken with this prompt, with this sort of checks having been done, because you will be audited or because you will need to do a review of something, something that went wrong in the past. And so to be able to see what AI did, what, when, in what context, and with what information will still be crucial. And of course, a lot of the work we do is collaborative, is collaborative between us and having you know a more traditional UX where you can see the output easily, you can see the plots, you can see the KPIs, you can see the tables, will still be important. And a way to share that among colleagues and to have discussions on a platform will still, of course, be important. I'm sure many of you have already started building agents. Sometimes building agents that really understand actuarial science is not simple. So what we at Akur8 are bringing to the market now is what we call expert actuarial AI agents. These are AI agents that are built by our actuarial team, by our product team. They are curated by them. And they will live within the Akur8 platform and will allow you to leverage all the execution layer that we already said through the large language model paradigm. The foundation of where our AI agents will live is the Akur8 technology, is an advanced and trusted deterministic technology. The AI agents and the underlying technology are curated by our actuarial team, our product team, so industry leading experts. And everything is transparent and traceable by design. Our agents live on top of a sort of pyramid of, you know, cloud infrastructure where you really are running everything. And the machine learning is built on top of that infrastructure, represents all the algorithms, but also all the institutional knowledge. And agents are just having access to all of that. And if they have access to all the execution layer, they become very powerful. They can do anything you can do within the platform as a human. And so the more features are present, the more abilities are present natively, the more agents have the power to do more. So Felix, over to you. FELIX: Thanks, Ludovico. So whether or not you're an existing Akur8 user, we run workshops for actuarial teams that are really dedicated to AI implementation, covering all the topics we just talked about. So feel free to reach out through our website if you are interested. We hope this conversation gave you a clearer view of what it takes to put AI to work in actuarial teams, from adapting processes to building trust and maintaining the right controls. Thanks for listening to the Akur8 Academy Podcast. Brought to you by Akur8. See you next time.
