Dr. Karolina Najdek: Welcome to our first webinar this year. My name is Karolina Najdek, and I work for DataArt from our lovely Munich office. With my colleagues Viktor and Dmytro, I will discuss a very interesting topic - the impact of generative AI on the insurance industry. So over to you, Viktor. How are you? How is the weather there?
Viktor Vasyukov: Not bad. Thank you, Karolina. Hi everyone. My name is Viktor. I'm a business analyst in financial practice at DataArt, specializing in insurance and insurtech projects. Within my daily duties, I'm really focused on understanding real business needs and opportunities and aligning that with what technology can offer to bring real value to our customers. So here today, with all of you, I'm really curious to have the discussion and explore what interesting things might be possible using generative AI for insurance. Thank you. Dmytro, let's continue.
Dmytro Baikov: Thank you, Viktor. Hi everyone. My name is Dmytro. I am leading our work from the technology side, and we have been doing a lot of things in generative AI for the last year. We are building prototypes and speaking with clients, so we have a pretty recent view of the industry and how insurance and financial services react. I am happy to share this knowledge with you and speak about generative AI and how to build these solutions today.
Dr. Karolina Najdek: Thank you very much, Dmytro. The most important thing everyone is probably wondering is what we will talk about. We definitely want to discover how generative AI tools are already reshaping the industry and providing solutions. We want to gain a deeper understanding, together with Viktor and Dmytro, of the typical workflow for implementing generative AI. And of course, explore a range of compelling use cases for this technology. It's super interesting. It was a very interesting 2023, and the year is not over. I'm pretty sure it's going to be a super exciting topic for 2024. But guys, what do you think about the impact of generative AI on industries in 2023?
Viktor Vasyukov: In my opinion, it's really growing now. During the last year, we've seen it evolve from a new emerging concept – really interesting, challenging, and sparking lots of conversations, questions of how it could be used, what is feasible, and whether it would totally change and disrupt our current ways of working.
The way I see it, during the last year we've seen it shifting towards the next phase of early adoption. And more and more companies around us are starting their journey, doing proof of concept pilot projects, first trying to get some real understanding of what value could be achieved practically, and then getting some hands-on knowledge on what it takes to bring it into daily operations. I mean such aspects as policy, accountability aspects, and similar issues. What does it look like for you and your colleagues?
Dmytro Baikov: I can tell from the technology side. Early this year, we saw companies starting to understand the impacts and generating use cases, but they were still very cautious about privacy and security. Throughout the year, we saw how policies changed, how all cloud providers evolved their compliance departments and privacy policies, and how open source models were deployed. So then we kind of shifted from privacy issues. We now understand really well how information is handled with cloud services.
They do not train models on your data. If you integrate business solutions and enterprise solutions, they don't need it - they don't want to ask clients about the product. The public version can be trained on your data, so it's not an enterprise service. It's very important to know and understand this distinction.
Now, the conversations are more like, "We tried this and that, we tried different POCs, now how can we move to production?" That's what we are seeing right now. This year's trends evolved from the idea stage, and now we see production integrations. From a technology standpoint, that's what I observe. Karolina, what do you think?
Dr. Karolina Najdek: What's very exciting for me, and I think about this every time I speak to insurers, is that they have collected an amazing wealth of data over the years. And what are they going to do with that? Because we all know - and I believe you, Viktor, from the business point of view, and you, Dmytro, from the technical point of view - if you have the right data, a lot of right data, you can do pretty amazing things in a company.
You can really support the insurance industry, don't you think? That's why I am really excited about the presentation you have for us today. And please, our lovely audience, do not forget to ask questions because this is such an exciting topic, and we will definitely continue to speak about it next year.
Viktor Vasyukov: Thank you for bringing up the data aspect. I think it's crucial here. I have an impression that there might be some kind of opportunity in what generative AI could bring to help conservative industries like financial and insurance.
I'll share what it looks like from my point of view, and please feel free to add more color to that since you're closer to technology and have much better expertise, Dmytro. The way I see it, within previous waves of artificial intelligence, it had more strengths with arrays of technical, machine-produced data. It had a very strong impact on industries such as manufacturing.
Now with generative AI, due to its different nature or approach, it gives a promise of removing barriers between knowledge workers - the people who need to digest and process lots of information already available in documents, human-readable text, and other sources, not machine-readable data. It gives an opportunity to bring all the data closer to the people who are at the core of such industries. Does that sound right to you?
Dmytro Baikov: I would say that historically, AI has had a couple of so-called winters. We couldn't train the models because there was just no data. Then the data started to grow. We started to get better processors, storage, etc., and now we have lots of data. With generative AI, we even have too much data because we can also generate data and train models.
So, from the technology side, we see that people use bigger models to generate training sets for smaller, more specific models, like industry-specific ones. You can take everything from GPT-type models and generate training sets for your private, classical machine learning models.
So basically, you can do all that. That's why the main question is how to leverage your data and generative data for your daily use. It's now possible to enrich the data. But I still think the main challenge is how to leverage your data, get insights, why do data analytics, and where to start with a use case. It's still the challenge, usually. What do you think, Karolina?
Dr. Karolina Najdek: Well, I'm still struggling when I speak to prospects and customers. When we go into discussions about data, I try to leverage how important it is to have the correct data in their systems because they are very confused when starting projects. They wonder, "Do we have everything? Do we have the correct data that we need for a project to be successful?" And I think that's what makes them very cautious at the beginning of a project. They know they have been collecting data for months, for years. But what's the quality of the data?
That's why I am very thankful that we have two points of view – business and technical – to see the difference in approach. In the end, we need both to have good results. We all speak about digital transformation in the insurance and finance industries. But what is that transformation about? This is what I believe many traditional conservative companies are afraid of.
Dmytro Baikov: Yeah, and I think that's the right time to jump to the use cases because you can plan for yourself as soon as you see how your competitors and neighbors use the data. It's not about stealing something – it's about how it can influence your business, how you can see the impact that is already there, and how you can apply it to your business.
So I would say it's about creativity and reusing from different industries. We work across different domains, and it's very common that something goes from retail to insurance and vice versa, as well as from finance. So, Viktor, it would be great if you could provide some insights on these use cases, and we can maybe think about implementing those or training models.
Viktor Vasyukov: It's a great point. I think it's a good time to look at several real-life success stories and examples that we've gathered to show that despite the technology being in the early adoption stage, there are already real results, some pretty impressive numbers in several insurance use cases.
So let's take a look at those. The first one is from auto insurance, which used generative AI to build an AI-powered virtual customer support system. With this generative chatbot, they were able to get a more than three times increase in the number of customer requests resolved right in place by the chatbot itself without any need for human support. At the same time, it decreased the workload of traditional email support by more than 50%. With the increased requests resolved right in place, they eliminated customer waiting time. So customers get their answers right away, which is significant for customer satisfaction and the feeling of interacting with the company.
The next one is related to the underwriting process. A specialty insurance company used generative AI to automate the underwriting process, processing submissions received from brokers. In this case, AI does technical work on enriching the data, getting additional data from different sources, triaging submissions, and filling in data in different systems. A lot of technical work takes up a big part of underwriters' time. Within this new process, expert underwriters start working with already prepared data. As a result, it gave more than a 100% increase in the productivity of underwriting workflows. It also allowed them to reduce turnaround for key strategic clients from 24 hours to just two, which is an amazing change in my opinion.
The next one concerns automated claims settlement. It is from a digital insurance company offering different lines of business. In this case, the system powered by generative AI fully resolved and paid out a claim in less than three seconds, a record-setting time.
Dr. Karolina Najdek: Oh, wow. That's really fast.
Viktor Vasyukov: Really impressive. Yes, we can discuss that this is a case for special conditions. But at the same time, it's a kind of benchmark that could be used as a standard. Working on putting more and more percentage of claims through this process with the same performance. So, as I said previously, even though we're in an early stage of adoption, we already have real cases with numbers and results that have been proven.
Dr. Karolina Najdek: I'm very interested in the claim settlement topic because it causes many insurance companies a lot of headaches. Viktor, do you think it's possible that generative AI can help automate this process completely? Is it possible at all? From my point of view, it's super complicated and connected to so many different data sources.
Viktor Vasyukov: From my point of view, we are talking about completely automating 100% fully, maybe eventually, somewhere in the future. Within the near to mid-term, I'd say that automation of such processes goes in the direction of triaging cases and separating them into standard ones, simple ones that could be streamlined, and the ones that should be referred to human intervention, processing them constantly. And then the question is getting a higher percentage of cases automated over time. That's how I see it.
Dr. Karolina Najdek: Absolutely. Dmytro, what do you think from a technical point of view? Your view is probably completely different from ours.
Dmytro Baikov: No, not really. Usually, these complex processes do not rely on technology alone. They rely on people, and sometimes, people are very resistant to changes in insurance, for instance. So maybe another insurance company is looking at that, and they say, "Well, it's something very small; it shouldn't work this way," etc.
It's great to see all of these companies moving to production and showing real benefits. The next level of chatbots—we already had chatbots, but now they are much smarter. They can perform actions like autonomous agents, helpdesk, or support personnel.
I would say it's a great leap forward, and we will see more solutions and products where, in a chat manner, you can resolve a lot of problems, issues, and requests. And the underwriting as well – if you can decrease processing time by 20 times, then increase the processing speed, we can boost productivity. I think that's great.
Maybe it's a small solution that gets some data out of a document. But on scale, it's a huge leap forward. We are building small, tiny solutions that help move 30-40 times faster, slower investments, but have a huge impact on the whole process if you have this process in place. So I would say it's great to see these things moving to production and improving their business metrics.
Dr. Karolina Najdek: It's very impressive. I'm really curious what Viktor has for us on top of that, which is already very exciting.
Viktor Vasyukov: I'd like to spend a couple of minutes talking about additional applications. So far, we've looked at just three separate use cases. But I'd like to mention that these are not the only possible applications for generative AI. Actually, there are lots of different ideas and potential applications. Here is a list of items allocated to different parts of the value chain, showing many potential use cases at different steps. This list doesn't try to be complete, but here are some examples, and there might be much more.
Generative AI looks promising and shows potential to help at very different stages. For example, it could start with customer engagement in producing personalized materials, interacting with clients, or even using virtual assistants or chatbots to help potential clients compare different products, select options, ask questions, get answers, and be assisted in selecting the right product for them.
Moving to the next steps, it could potentially be really helpful in client-facing operations. The same customer support could be external chatbots that clients use themselves. It could be internal assistants bringing agents and employees all the needed information, personalized for specific interactions, and bringing the power of the knowledge base and huge loads of information that the company already possesses.
It could be used to automate pieces of back-office operations, such as underwriting and claims settlement as we briefly touched on, as well as detecting fraud in claim settlement. However, there might be many more uses, and even such processes and business functions as product enhancement and long-term insurance improvements.
There could be examples of using generative AI to do risk simulations, generate different risk scenarios, premium calculations, and modeling. It could also be used as an aid in generating filing documentation, which might be a really expensive process requiring significant workflow. Some technical parts could be taken over by generative AI. I'd say the next step is further exploration to see how to design the next steps for these directions. What do you think?
Dmytro Baikov: Yeah, I think it's important to understand how to do it. So maybe we can walk through how it will be done and how we will do it in detail on the next slide. Usually, all of these use cases can be split into phases. Of course, you start by defining your AI journey and use case, and start by understanding your business. You build hypotheses. Then you can select one, two, or three priority use cases. So the first part is ideation.
But as soon as everything is set up, we have these four phases. We build on top of the cloud, we build on top of open source, and we start from prototyping. It's very important to prototype to see the first results, research the data, see its quality, try different approaches, libraries, and algorithms, and evaluate the results. Then, we create a feedback loop to enhance it until we have the first clickable prototype.
As soon as you have this first clickable prototype – it may be an API or a simple UI – you see the impact, how it works, what outputs it gives, where it's working, and where it's not working. So, you become much more confident in building AI solutions overall. You see how it works, and then you can scale it, move it to the cloud, or deploy it. This is what we call the MVP stage – MVP development.
You can proceed to fine-tune the model, to enhance the data and approaches. But still, you will have the model deployed. You will have the first feedback. You may integrate with other systems - add more storage, logging, and monitoring. When you go live, you will collect feedback from real users. You will see how it works not only for you but for your people - it may be internal people or external clients. You will optimize based on all the knowledge and feedback, and may proceed with A/B testing or model enhancement. But now it will become part of your product and features.
The support stage may be different. You can add minor enhancements. If the data changes, you can assess the effectiveness. If it's working too slowly, you can consider different models in the future. But this is the usual project flow, and starting properly is very important. This way, we have these four phases to make it to your MVP and deployment to production.
So that's how we build. You can explore the ideas that Viktor mentioned. If you want to repeat something that is already there, this is our suggested flow, which we find the most success-oriented and impactful for our clients. We really see the impact when we prototype and move to MVP, rather than saying "let's build, build, build" as a full project. We build it in phases using an agile approach. So, Viktor, what do you think about this flow? Does it make sense to you from the business side?
Viktor Vasyukov: It looks great and really is the way forward, providing further knowledge on what exactly lies underneath all of these ideas. I'm really excited to see the next steps.
Dmytro Baikov: Yeah, that's great. And actually, it's because when you build one use case, you probably want to build a second one. That's how it works. You will find more and more uses. We call this the AI solution or generative AI solution. It's one piece of functionality - it may be a feature or an application. And it's deployed. So now you're open to testing your ideas and use cases.
Of course, when you have the second and third ones, you follow the same flow. And now you want to be able to add faster than 6 to 8 weeks. You want to build and deploy to production faster. Then we go to the platform level.
We see that it's very important to have a platform component when building scale solutions. You have the data pipelines, the data engineering, the scalability, the monitoring, and the cloud tools for building all these use cases. So, you will have one single platform implemented and integrated with all the use cases. Now, you focus both on what's in production and making the first iterations on prototypes in the early stages of your solution.
Of course, when the platform is right, it's time to extend even more and go to the factory level. If you get one document processing solution for one department, you can probably reuse the same approach with different data for another department. That's where the use case scalability part comes in, and you will have some use cases that are already working.
If you're a big company, you want to scale to different people, products, and internal and external services. So you will have blueprints that you can reuse. That's what we call a factory. If you have a working piece of functionality that you can reuse with different data or different cases, it accelerates this R&D work a lot while still keeping the AI platform component for production usage.
You may take this journey from the technology side in the AI world. It can be implemented differently on different platforms. But that's how we see clients may evolve from the technology perspective. I think it's a pretty good view. You may have a different one, but that's what we see.
Dr. Karolina Najdek: It's super interesting. Before we move to a Q&A session in a few minutes, guys, if you look at the whole year 2023 and think about 2024, what do you think we should definitely think about? What should we plan, or what should insurers consider for the coming year?
Viktor Vasyukov: That's a good one. Let me start. Well, about some predictions and things I've heard around, there is an expectation that within the next year, there will be more and more employees in companies using generative AI in their daily work. That might be something we need to face and get prepared for. It's moving to certain stages of adoption, whether we like it or not. The prediction is that it's going to happen this way.
Another prediction – and it's hard to say about the exact timeframe – but there are expectations that with the development of generative AI, it should become more powerful in terms of creativity and problem-solving. Going to the next stage, I think that might unlock new qualities and even more use cases. What does it look like from the technology point of view, Dmytro?
Dmytro Baikov: Yeah, well, I think the way to go right now is toward different modalities, as it's called. So it's not only about text, but about images and video. If you look at the new model released by Google, Gemini works with different inputs - it may be voice, video, or images.
So, I think next year will help us build these types of solutions more easily, like image-to-text, text-to-image, and more. It is very important for companies to evaluate this because it will enhance document processing and chatbots, as you'll have better models. I think next year will open a broader range of the same solutions we have now. So we'll have chatbots with even broader functionality. I think it will evolve in the assistant space, where AI tools can do four times more with integrations and with these multimodal tools.
That's what you can start thinking about right now. If you have, for instance, text and images, and you can currently work only with text, you can already plan what you could do with the images from the ideation step. So I think those will be what we'll see more of - more data inside the models and more data to use with the models.
I think it's probably the right time for a Q&A. I also want to promote our workshop, which we do at DataArt. You can scan the QR codes, and if you're interested in discussing how generative AI can work in your company, what tools you can use, how to start, and how to move to production - that's what we do. So feel free to send a request, and we can chat about that. Otherwise, Karolina, I think there are some questions.
Dr. Karolina Najdek: Oh, yes. There is a lovely question. When I look at it, I think I could write out my statistics about it, but this one is definitely going to Viktor: "How do you foresee the evolution of AI-powered virtual assistants or chatbots in the insurance industry to deliver personalized and empathetic customer service experiences? What strategies can insurance apply to ensure the ethical use of AI in customer interaction?" This is a very complex and lovely question. Thank you.
Viktor Vasyukov: Several questions within one. Let me try to start with this one. Within chatbots, I'd say that some aspects need to be set. I think that with chatbots' potential growth, there might be an ability to process the emotional part of communication with the customer, scan the condition and emotional state, quickly pull in a human if needed, or provide better-tailored, better-shaped responses.
Regarding ethical aspects, I believe that's a huge question in itself. Some thoughts regarding strategies: I believe strategies should be built into responsible policies, followed by policies, and elaborated on during implementation.
As I said multiple times today, we're in a stage of early adoption. This means that approaches and patterns are still emergent, and regulations and recommendations are also emerging. I believe what already exists within different countries is more like recommendations. There are very few, if any, restrictive regulations apart from previous data protection laws like GDPR.
In addition, there are various potential ethical challenges. Generative AI tends to reproduce bias from the training data - again, we're speaking about the quality of data. And it may deteriorate when "fantasizing." I definitely will ask Dmytro to add additional color with his more practical expertise.
But I believe there should be policies at the pilot and proof of concept stages – monitoring the output of generative AI, fine-tuning, tweaking, and making sure that what it produces, the outputs, and results, are in line with the company's policies.
Dr. Karolina Najdek: Thank you, Viktor. Amazing. This is such a great question.
Dmytro Baikov: Yeah, I can also make a small comment on this one. About the empathetic customer experience - I think it might be even more empathetic than the average person, because if it's a long conversation, robots won't be slower or ruder with the correct settings, unlike humans sometimes. So I think if its only goal is to serve customers better, it will do the job, and that's how it's designed. That may not be the case for a person - they might go for coffee or forget something. So I would say that, on average, the customer service experience should improve.
Regarding ethical usage, many developers are building guardrails and tools to help you stay safe. It's called responsible AI. We are also implementing techniques on the prompt engineering side, which can keep you on a single topic and not offend anyone. However, the question is more complex for companies developing these AI models.
I would say that developers mostly rely on the creators of these models. Since you do not often train these models by yourself, you just fine-tune them. Of course, you can fine-tune them with malicious intent, but that's still how it works right now.
Dr. Karolina Najdek: Thank you, Dmytro. I have another question for you: "What skills and competencies are essential for insurance professionals to work effectively with generative AI?"
Dmytro Baikov: Yeah, I think it's a question of personal productivity overall. Humanity should understand what a prompt is and how to write prompts. It's all about correctly expressing your thoughts so that machines can understand. You need to understand some rules, but overall, if you manage to do that, it will be like a dialogue with the machine where you can be confident in what you ask and what it replies to.
So I would say prompt engineering is a unique skill, number one. And the second skill is to try many productivity tools that already have integrated generative AI features, understand how they work and how they help, and experiment with them. As soon as you see that, for instance, GitHub Copilot is working really well for coding, that's the way to go. The same for Notion or other tools.
So, if you can include tickets in Notion or Jira, that's great. This would be your skill – not being afraid to try these new tools, even if they aren't working perfectly at first. Continue and try to find your use case. So, I would say these are the key skills.
Dr. Karolina Najdek: I think we still have time for one more question. This one is super interesting as well. So, thank you to the audience for asking amazing questions, because this is what this webinar is about - organic discussion. So, guys, you decide who answers because I think it's for both of you to be honest: "What strategies and practices should insurance companies adopt to ensure compliance with regulations and standards such as GDPR, SOC 2, HIPAA, and ISO as they integrate generative AI into their operations?" Big names everyone is scared of. So what do you think?
Viktor Vasyukov: It's a big question. I'll try to start with a couple of thoughts. I believe you should use a privacy-by-design approach and all the practices related to data loss prevention, assessment of data protection impact, and so on at the stages of use case and concept design. So, having all those principles and approaches applied from the very start in this context. Dmytro, please share your view on this.
Dmytro Baikov: Yeah, I would say that you need to be aware of the new developments as well, like the AI Act from European governments. Something will probably appear from the US side as well, so you need to understand how the government supports it overall.
Some clients and cloud providers are hyper-compliant in general, so you can check the documentation and privacy policies of your tools. Regarding GDPR, I think it's basically the same – you need to understand how you store and process the data, and generative AI falls under processing.
I'm sure there will also be some different rules for ISO, so please check how it may impact your systems – it may not impact them at all. My recommendation would be to take a look at the security policies as well, not only for data governance but also for securing your solutions. This is something that is important with GDPR as well. So this is my take.
Dr. Karolina Najdek: Absolutely. Thank you. Guys, this is so interesting. I could speak to you for the next two hours, but as you can imagine, we all still have much to do before Christmas. So I would like to remind everyone listening and watching not to forget to scan the code and book a generative AI workshop. You will not regret it. It's great fun. You will learn a lot, and this will definitely help your company in digital transformation and innovation next year.
Thank you, Viktor. Thank you, Dmytro, for your time and all the amazing use cases. I hope we see each other at the next webinar, and next year we'll see how the future of Gen AI is progressing in the insurance industry.
Dmytro Baikov: Thank you, Karolina, for great moderation and leadership. Have a great day, everyone.
