This guest blog post was written by Andreas Engler, the former Director of Engineering at Nexible, a German insurance start-up that’s been using RapidMiner to digitally transform the way that the insurance industry does business. In this post, he provides five key use cases that are particularly beneficial to insurers, based on his own experience.
Many of today’s insurance companies are stuck with out-of-date bureaucratic processes and are in nearly the same position they were five years ago with regard to using artificial intelligence and machine learning—which is to say, they haven’t done much with it to impact their processes. Even many major insurance companies have publicly announced that their information-driven business model is ripe for enhancement with the use of AI. “Data labs” have been implemented all over the industry—and yet, very little machine learning has been implemented into the core processes of these businesses because they’re stuck with old ways of thinking—this is how it’s always been, so this is how we do it.
But that doesn’t mean that digital transformation is out of the picture for insurance companies. A closer look at the business model for insurance companies confirms that data-driven decision-making, powered by more robust systematic data collection, will enhance insurance companies’ bottom lines.
The business model—which consists mainly of predicting individual customer risk better than the competition and operating the most cost-efficient claim-coverage and customer-support processes—is an optimal use case for machine learning. It thus makes sense that applying AI and ML to insurance industry challenges will introduce strong economic benefits.
In this post, we’ll look at five of the key ways that insurance companies can use AI to impact their business model.
Five Ways AI is Impacting the Insurance Industry Today
It’s undeniable that improving the usage of artificial intelligence can produce benefits in key areas of insurance businesses.
Over the past four years, I’ve helped guide Nexible, a German insurance start-up, through the process of adopting AI and ML into the technical platform for various use cases, including a support AI, a chat-bot, an automated claims-handling system, and adapting more advanced tariff-model algorithms with a real-time pricing engine.
This combination of tools provides Nexible customers, third-party claimants, and investors with significant benefits. Let’s take a look at five areas where these tools have had a substantial impact on Nexible’s bottom line.
1. Insurance product design
A more purposeful collection of customer, operations, and claim data paired with advanced data analytics techniques like machine learning leads to better insurance products. These better products, in turn, cost less to be advertised, convert more potential customers to contracts, and find customers who are less prone to churn.
Establishing a customer centric data collection process provides an easy opportunity to analyze similarities and lookalikes regarding campaign tracking in the sales funnel (and later) by looking at all of the operational and claim costs for different customer groups. This gets you into a position to build classification models to predict future sales performance for new products, as well as build ad-campaigns for different channels that target different personas based on your own data, rather than relying only on what Google and Facebook provide.
2. Automating operational services
New ML classification algorithms allow for the automation of operational processes, moving beyond classical, hard-coded automation that has a very limited scope to something that is able to handle complex tasks like those of support and claims agents. This introduces the opportunity to provide faster, higher quality, and less expensive services to customers and claiming third parties for all channel interactions.
For example, Nexible’s support AI handles unstructured emails and presents customers with an end-to-end process automation that typically takes less than 15 minutes for most customer service intents, including things like coverage changes or moving to a new address. Automating these straightforward tasks helps to keep operational costs low.
To create this support AI, Nexible agent teams manually tagged customer email intents for several months in the ticketing system. The classification models were then trained and validated based on the information available in the ticketing system.
With the models in production, the AI service presents itself as a first-level agent handling tasks, working in parallel to the agent teams. This way, the agent teams can control what the AI service is doing and can provide input about how it is doing, creating a feedback loop to continuously improve the service.
By using RapidMiner Studio and AI Hub, the modeling and productive usage of the models for the important first six intents took less than 120 hours for the complete project (excluding the manual classification beforehand).
This project has resulted in a significantly better customer experience, reduced operational costs, and a huge increase in trust in AI-driven automation solutions within the company. This kind of cross-department buy-in helped to create momentum for other ML projects. Similarly impactful in this regard was the detection of relevant customer attributes discussed above, which demonstrated the power of AI and ML by significantly impacting claims risk and operational effort.
3. Actuary modelling
Newer algorithms and more data also allow for the enhancement of pricing predictions—as well as the individualization of prices—for different customer segments. This provides an edge against competitors and helps attract the best customers in terms of claim and operational costs.
Good actuary departments understand that machine learning algorithms are not competition for their work, but instead are a new tool to help them do their work more quickly and effectively in comparison to a simpler model approach. Additionally, tools like RapidMiner can provide a higher standard of automation, verification, and validation in the context of the work done by the actuary departments.
4. Claims services
Using AI for our claims service has been more challenging, legally speaking, than the projects discussed so far. As the service is a combination of classification and regression regarding the claims’ coverage and the claims’ cost prediction, it has the potential of denying the claiming party a certain amount of money. In Germany, such a decision is legally prohibited to be taken automatically. Therefore, manually interaction needs to be included in the decision-making process even if the AI model’s predictions would be as good or even better than human decisions.
Additionally, from a statistical perspective, claim cost regression is a challenge for a number of reasons. For example, repair costs at different German repair shops are highly variable, due not only to different car models and damages, but also due to the local area and legal status of the repair shop itself, creating a large number of variables. Here, the introduction of repair cost brackets can improve the prediction quality significantly and allow the implementation of a semi-automatic process in which the customer can get a fast payout option automatically, with each subsequent claim rejection or interaction done manually until further models can be trained.
Depending on the insurance product, this challenge is most likely to occur in a multi-dimensional space where the data points of just one insurer are not dense enough for good predictions for car repairs. Here I would recommend that you partner up with external vendors or build a community with other insurance companies to share data in order to improve your models and their predictions.
5. Fraud detection
One of the most interesting and challenging topics regarding AI in the insurance space is fraud detection. Finding patterns of fraud and improving the quality of detection in comparison to expert handling, while also automating the task itself, has the potential to improve both claims costs and operational costs.
While the later can be achieved via classification models in the operational processes and therefore presents itself as a similar challenge to any other automated process, the classification of claims as fraud historically lacks verification, as not all fraud cases are validated for fraud after not paying the claimant. Instead, in most fraud-labeled cases, if the claimant gets no money or service, the claimant doesn’t have many options to demonstrate that the claim wasn’t fraudulent besides going to court. Because the final true answer isn’t ever known, this bias in fraud handling would simply be passed over to any classification model trained on the data.
Don’t get me wrong, fraud management is a needed element in the insurance sector. However, best practices in regards of validation cannot be applied for fraud cases as the actual state of fraud will always be unknown to the insurer.
Additionally, AI can be used to deliver new scores and provide added legitimacy of a claim by, for example, examining the authenticity of photos, the similarity to previous claims by other customers, or even similarities to all of the claims in the last year. This area represents a big opportunity for companies to provide insights for the insurance sector in a certain country. Therefore, in my point of view, a partnership is recommended around fraud even if you manage one of the biggest insurance companies in your region.
Why Isn’t AI Adoption More Widespread in the Insurance Industry?
It’s clear from these five examples that AI has the potential to have a significant positive impact on insurance companies’ bottom lines. Then the question becomes why newer and more advanced data science techniques have not already been broadly adopted in the insurance industry as they have in, for example, the e-commerce sector.
In my personal experience, the lack of adoption is closely related to two things: an insurance-industry working culture that is in most cases still not very agile, paired with legacy technology that can’t easily be adopted and/or interconnected with data science best practice data-collection and runtime environments.
When we started our journey to adopt AI at Nexible, we were fortunate that neither of these hindrances where present, which meant that we could easily adopt advanced machine learning techniques as part of our core business processes from the start. If you have this advantage in your company, you should embrace it and start preparing the architecture and all of the operational processes that you’ll need in the future.
If you’re not so fortunate, you need to start thinking about establishing an agile culture for all the projects you’re working on, not just AI projects. Additionally, your data science team needs to be decentralized to incorporate all relevant business departments. Data scientists are more like consultants, broadly adapting to the needed business cases in order to find and implement new and innovative data-driven solutions for your problems—so long as you provide them with the needed environment of a RapidMiner and/or Python stack and access to the relevant data and legacy systems.
Consider this too: every data-driven technical advocate in your company with a good understanding of the processes and the related business value behind the use cases can become a data scientist if you provide them with the right tools. Applications like RapidMiner mean you don’t need experienced Python-coding chops in data science to leverage machine learning and generate business value in your company.
Looking Ahead: What Does the Future Hold?
Looking ahead, the legislative and regulatory environment will most likely prevail over this decade well into the next one, at least for the EU region, which will somewhat limit the adoption of AI tools by insurance companies. However, tools like RapidMiner will continue to grow in their ability to enable companies to easily adopt more advanced artificial intelligence techniques into the legacy tech-stack present in the insurance sector.
Therefore, the biggest challenge for the industry that I see will remain their complex business processes and non-agile working culture. However, as many insurance start-ups and established companies have already demonstrated, these challenges can be overcome, leading to broad adoption of AI and a digital transformation that creates a compelling competitive advantage.
At some point, this competitive advantage will not be there anymore. Instead, a disadvantageous market position and declining business will be the results of non-adaptation of AI and ML as companies that operate on the older model fall behind their competitors.
While the principal use cases for ML in insurance won’t change in the coming years, the extent of adaption of these advanced tools will. And the impacts from AI and ML won’t be limited to insurance coverage and pricing but will also include the overall UX and customer interaction even more than it does today. Here, AI-driven product features will ensure a more convenient customer experience and higher customer retention. While today’s automation use cases focus mostly on the agent workforce, the automation in decision-making will extend to other business functions present in the insurance sector, like business development, marketing strategy, mergers and acquisitions, and even management itself.
However, as nearly every country provides an extensive regulatory body for the insurance sector, the amount of AI adoption will always be less in the insurance industry in comparison to non-regulated sectors. But that doesn’t mean that you shouldn’t use it as much as you’re able to.
If you’d like more information about getting started on a machine learning project of your own, check out our whitepaper A Human’s Guide to Machine Learning Projects to make sure you’re doing things right from step one.