By the looks of it, GenAI offers almost unlimited options. It made us think of possible applications in the insurance industry, more specifically: the vehicle insurance industry. We experienced the bureaucratic insurance processes first hand and decided to develop an alternative capitalizing on the current customer value chain: quick, accurate, cheap and fraud-resistant GenAI.
Initially, my role focused on processing the second and fourth lecture, which contained relevant definitions and methods I wanted to implement. A key aspect was the business model, specifically: for whom are we creating value? We opted for a broad approach designed to benefit both the insurer and policyholder. Capturing value was also on the agenda. I assessed that the advantages of our AI solution for the insurer were stronger, compared to those for the policyholder. I thought that the more policyholders would know about ClaimAI, it would become even more appealing for insurers. I proposed to monetize the service through insurers, effectively subsidizing the policyholders. This structure enabled us to create two-sided network effects. Applying these theoretical methods from the lecture to our solution was interesting and rewarding to do. In very concrete terms, I worked on the information we needed in slide 5, 6 and 7: business model, costs and pricing.
Almost immediately we came up with the ClaimAI idea. This led directly to our most important and strategic decisions: transforming an existing company or building a new venture. We ultimately decided to start a new venture. This track provided the best opportunities for creative solutions, because we could start from scratch. On top of that, creating something new allowed us to formulate a strong business case that would clearly deliver value to all stakeholders involved. Following our initial concepts, we created an iterative process by checking and redefining our options through (peer) discussions. Later during the prototype testing-phase, this iterative approach was repeated.
I was particularly surprised by the first prototype developed by my peer. It was impressive, which I told him directly. Looking at the prototype’s development after we jointly formulated and implemented feedback was a surprisingly rewarding experience. Obviously, we were confronted with some failures during the process. For example, the system failed to deliver on its promises if a claim just under €1,500 was filed. This was the auto-settlement limit and required countermeasures. A specific countermeasure was designed regarding the “someone was injured’’ box. After implementing the measure, any detected health cue would instantly abort the €1,500 auto-settlement. A human adjuster would interrupt the AI cycle at that point.
Feedback and empirical evidence remained critical during the process. Earlier I explained how feedback played an important part in creating an iterative cycle in the early stages of this project. A crucial moment during the project unfolded when we started with evidence-based testing. During evidence-based validation, we established performance thresholds. These thresholds were objective pass-marks to measure the quality of the model, and were an improvement compared to the initial subjective thresholds we planned to use. Repair costs estimations would only be acceptable if they fell within 30% of the actual repair cost the policyholder would communicate through the application. On top of that, the system had to detect at least 80% of the damaged parts. Benchmarks of other multimodal models showed 70% to 76% accuracy. When the initial test revealed significant variance in estimated repair costs, we added a fixed tariff matrix to our Generative AI calculation engine. Consequently, evidence obtained from testing directly influenced the direction of our solution.
Over time and when it comes to collaboration in a group assignment, I learned that it often presents challenges. In previous academic projects, I worked alongside peers who were unmotivated. I noticed that these settings negatively impacted my own drive, which is obviously a pity. However, this time I found myself in a motivated group with smart and creative peers. I enjoyed our collaboration and I value their participation a lot. During our time together, in terms of hard skills, I learned more about developing the model. This was a completely new and rewarding experience to me. Regarding soft skills, I got reminded again of the importance of clear and proactive communication in our interdependent workflow. Because a peer’s progress depended directly on parts I was still finishing, clear coordination was necessary. By sharing notes, core ideas and intentions top-down before the work was 100% complete, we ensured that the next steps could proceed without delay.
Prior to this project, I viewed Generative AI as a disruptive force that could be implemented in a plug-and-play construction. However, the project changed my perspective: I now understand AI as a strategy using a more holistic approach. This holistic approach requires looking at the whole organization, refining the data and most importantly, always paying attention to customer needs.
While Generative AI drastically lowers the cost of prediction (Agrawal, Gans & Goldfarb, 2018) (Lecture 6) and combined with algorithm performances, successful implementation cannot solely rely on these two aspects. During the courses and project and thanks to the theoretical frameworks, I gained insight into the important role human judgement still has. Either through evaluating the cost of errors or through designing the Generative AI in a way that consequential managerial actions can be taken.
Ultimately, sustainable value creation depends on the way an organization processes customer needs in the product. That is why deploying a public AI model barely returns competitiveness, as per the VRIO framework. Our struggles with repair costs variance and the €1,500 settlement boundary showed that sustainable advantage required embedding the AI with our own domain rules.
Moving forward, this project has demonstrated that AI alone is not a business strategy or a business case. AI is an enabler whose value depends on the design of an organization and how effectively humans collaborate with it.
