„Artificial intelligence helps close any holes in digitalization in order to cover all processes in a single workflow. There is no alternative to this."
Read the interview with Jan Langkau, division manager for the drebis and Aisaac products at adesso insurance solutions GmbH, to find out why artificial intelligence is so important for insurance companies and what should be considered when selecting a provider.
Staff: Mr. Langkau, what do you believe is the main problem when digitalizing processes?
Jan Langkau: The insurance industry is very complex. Unfortunately, this complexity often leads to processes and media being digitalized with relatively little reflection. The insurance companies expect more efficient and slimmer processes and therefore potentials for savings when implementing new digital systems. Nevertheless, „blind spots“arise again and again; so-called digitalization holes that are simply overlooked when developing the strategy or have been classified as subordinate.
Staff: What are the consequences of these holes in digitalization?
Jan Langkau: The holes in digitalization have two negative effects. On the one hand, all savings potentials are not used that result from consistent digitalization. On the other hand, there is the risk that the employees experience digitalization as an abstract mammoth project, but they do not feel any, or at least any positive, effects in their direct working environment.
Staff: That sounds very abstract, what does a process like this look like in practice?
Jan Langkau: Despite all digitalization efforts, insurance companies still have to handle thousands of paper documents every day. Employees view, classify and edit emails, letters and faxes. Media discontinuities occur here, which cost time and money. Because despite the use of digital systems within the company infrastructure, the consultants do not feel an improvement in their work situation. They are exactly in this digitalization hole. And exactly here is where there are enormous potentials for savings and efficiency. The conversion of incoming unstructured documents into digitally structured and therefore machine-readable data really utilizes the potential of digital systems.
Staff: And how do you digitalize documents? What steps are needed and when can you start using artificial intelligence?
Jan Langkau: The first step on the way to the automation of incoming documents is the digitalization of content through optical character recognition" (OCR). The process can be technically divided into two phases. Upstream image optimization that works similar to the camera software found on a smartphone. It automatically corrects the image resolution and orientation of the documents. The optimized output material runs through the OCR in the second phase. This creates a digital equivalent from the printed letters, in the form of deep neural networks thanks to artificial intelligence. OCR technology has made astounding progress over the past few years. The errors that sometimes still occur in this phase, for example, are corrected by special dictionaries. Image optimization and OCR are nevertheless CPU-intensive processes. Service providers are confronted with the challenge of ensuring performance throughout the entire process.
Staff: Is that all there is with using artificial intelligence? That isn't really anything new.
Jan Langkau: No, the actual heart of the digitalization of incoming documents is hidden in the next step, which is not conceivable without AI. The system develops an understanding of the content of the respective document without which the further automated processing would not be possible. Fixed rules and AI have to be combined with each other to classify content and extract information. An invoice, for example, can be determined based on certain properties even without AI. CPU-intensive artificial intelligence is not needed for this. But anywhere that clear rules cannot be formulated and criteria is only available in a subtle manner, algorithms from machine learning are used.
Staff: How can a machine or an algorithm learn and understand what the right answer to a question is?
Jan Langkau: With machine learning, learning algorithms analyze a representative data quantity based on its components, meaning words or word pairs. Then the most important words or word pairs are identified and linked by a defined question-answer combination. This process is frequently repeated with different question-answer combinations and different words or word pairs. The algorithm learns this way and becomes an expert for making a classification decision. The AI training, however, requires human support in the beginning, because the selection of the data which the machine should use to learn is performed by the people.
Staff: How well does the much vaunted "off the shelf" solution work?
Jan Langkau: My experiences have showed me that the customers are ultimately unsatisfied with an "off the shelf" solution. No document is like the other and each industry has its own special and subtle detection properties. That is why the application should be developed as customized as possible for the concrete application so that nothing stands in the way of enjoying the product in the long run.
Staff: There are a lot of providers on the market. How do I find the right partner?
Jan Langkau: The solution providers are different from each other in the quality and architecture of the machine learning. That is why insurance companies should allow the provider to explain their individual approach. Modern and contemporary AI systems combine different procedures with each other to combine the strengths of the different learning algorithms. This is how different virtual experts decide together on how an object should be classified. Such an ensemble approach offers the highest hit quality and requires corresponding experience from the provider in the implementation and processing of diverse document types – from texts containing a lot of context to table formats such as invoices. This in turn requires correspondingly comprehensively trained AI models.
Staff: Are there other parameters that should be paid attention to when selecting a provider?
Jan Langkau: As an insurance company, you should have them show the interaction between human and machine, at precisely two points: on the one hand, when the learning algorithms are configured in the beginning and on the other hand, when the consultant acts as the "teacher" for the machine during their quality assurance work. This is the only way that the actual scalability can be assessed. You should also be able to change processes and document types at any time to permit a modern configuration solution during running operations. There should also be no limit for the documents to be digitalized. This isn't needed either, because the machine learns quickly and, in the end, will almost do all the work itself sometime. My next recommendation is related to this. Sometime in the near future, the AI machine learning unit should take over the greatest possible amount of work. If human help is needed permanently, that is too expensive and barely scalable. If you are looking to invest in an AI-driven IT solution, you should take a closer look at the manufacturers and ask about it.
Staff: What comes after the digitalization holes are closed?
Jan Langkau: The next logical step would be "Predictive Analytics". This allows AI systems to use existing data to make concrete predictions about the development of different problems. Through the analysis of results, they offer support in the recognition of potential risks.
Staff: Thank you for this extensive interview
Jan Langkau: No problem, any time.
Do you want to find out more about our AI solutions to close your digitalization holes? Then contact our expert Jan Langkau directly!