Underwriting Insights: Automation Accelerates the Process
The underwriting of large risks is highly complex. With automation technology, (re)insurers can unlock insights in the data and profoundly improve their risk-assessment capabilities. One of technology’s biggest advantages is the intelligent delivery of pertinent information to the right underwriter at the right time.
Since the advent of the digital revolution, the banking industry has shown considerable improvement in understanding the needs of customers and adapting to them. The creation of data and the ability to extract insights from the data have proven to be powerful tools in this regard.
Upon receiving a reinsurance submission, an underwriting assistant manually enters the data into a workflow system for costing and quote preparation. This involves reading the entire document, which typically ranges from fifteen to fifty pages, and locating key data attributes which describe the risk. These attributes are used to create models and supply the underwriter with information to decide on whether or not to take on the risk.
However, this baseline dataset is not enough to fully illuminate the complexity and nuances which must be considered in underwriting. Information-gathering about similar cases and their costing is just as cumbersome and time-consuming.
The vast majority of data stored is unstructured. In other words, data without a defined data model, such as text documents, e-mails and pictures. Large-scale analysis of this data relies on more better tools than what most companies currently have on-hand.
The world of submissions is equally unstructured. These files often take the form of either a Word or PDF document with minimal structure and considerable complexity. There are clauses and exclusions which determine the parameters of the deal. To add to the complexity, many of the PDF files are scanned. This requires more time to convert the image to text.
Unfortunately, the industry does not follow standards in this regard, and most of the submission documents vary between clients and year. It is therefore difficult to model this space and build fully automated solutions. In addition, the dataset is relatively sparse. The number of annual submissions received by a midsize reinsurer is nowhere close to the number required to make use of artificial intelligence (AI) solutions. An AI solution typically requires at least hundreds of thou-sands of submissions.
This is just one of the many datasets relevant to the process. When considering the full scope of relevant information (e.g. news, internal research, e-mails, conversations), it becomes clear how much more work is required to illuminate the risk landscape for an individual submission. What’s missing in the industry to have this landscape be drawn automatically and to a much larger extent than it is today. Imagine clicking on a digital contract and having your screen filled with the most pertinent conversations, reports and similar past submissions.
Wrangling the data
The first step towards realizing a unified and dynamic risk view with unstructured data is through tagging – or the process of structuring the unstructured. A news article might receive the tags «negative», «cyber security» and «New York City» by applying sentiment analysis, topic modeling and entity extraction, respectively. Once tagged, this data can now be filtered and used for broader visualizations and data aggregations. Questions arise like: «Which topics and respective frequencies are represented in the news today?» «How does this differ between cities?»
The range of complexity for producing this structure is vast. One can extract keywords and known entities like client names, contract IDs and contacts with relative ease. Things become more complex when we enter the territory of neural networks and other machine-learning methods for extracting sentiment or classifying parts of text as relating to certain topics.
By processing an incoming submission in similar fashion, it becomes easier to issue figurative commands like, «Show me all news which is relevant to this submission.» In this case, «relevant» can mean shared topics, entities and so on.
The way forward
The future of underwriting is not without humans. This is a process far too «sophisticated» to be a contender for full automation, and one should be highly skeptical of any claims to the contrary.
Instead, the future of underwriting is an exercise in applying augmented intelligence, that is the integration of technology into human decision-making. It is the job of a machine to know which information is the most relevant to an underwriter. And it is the job of the underwriter to make a decision to take on or reject the risk. One can compare the situation to surgery, where the medical assistant – i.e. the machine –hands the surgeon – i.e. the underwriter – the right tools at the right time.
In order for this future to be realized, an augmented intelligence solution must include the following elements:
No SQL databases
The vast majority of the processing and querying will be with unstructured text data. As such, the appropriate database solution is required. Traditional SQL databases will not handle this well. Options include «MongoDB», «Elasticsearch» and «Solr».
The process of structuring the relevant scope of unstructured content is, in part, a natural language processing (NLP) task. Locating entities like cities, people and companies requires a solution with NLP capabilities.
Enterprise-specific mashine learing models
For more complex classification such as topic modeling, sentiment analysis and clause identification there is no one-size-fits-all solution. The best machine learning solutions involve large volumes of internal data to train models specific to an organization’s challenges and needs.
Given the array of problems that must be solved to craft a valuable augmentation solution, it is unlikely that a single application will fulfill all needs. As such, a solution must be pluggable. This allows other services, tools and applications to be integrated and do whatever they do best. The modern approach to integration utilizes RESTful API services1 for seamless communication.
Realizing this triage solution requires companies to take on new technologies and tools and bringing them together for a single harmonious user experience. Companies should expand their search for technology beyond large vendors (e.g. Oracle, Microsoft and SAP) to smaller technology companies.
Benefits of augmented intelligence
These elements will enable an underwriting triage process that has of yet not been realized. The benefits of this approach include, but are not limited to:
By presenting underwriters with all relevant information from important sources, discovery/research time will be cut dramatically and allow the underwriter to focus on the most promising submissions.
Improved risk management
As far as we know, the human brain is the most complex entity in the universe. But it sometimes misses important details. A machine-driven approach assessing risk information – like related previous contracts, noteworthy news and snippets from internal call notes – will dramatically improve the underwriter’s ability to distill massive amounts of information into insights for decision-making.
Elevated customer service
By combining the two benefits above, we inevitably arrive at a company able to turn over more business with greater confidence. This results in smoother customer interaction and better performance.
The future of underwriting will involve a blend of people and technology. One of technology’s biggest advantages in this space is the intelligent delivery of pertinent information to the right underwriter at the right time. Since this information is scattered throughout unstructured datasets – like news, contracts, e-mails and research reports – a technological solution is needed to transform unstructured noise into meaningful insights. The upfront investment will be well worth the reward, and we will soon find ourselves in a world where such capabilities define which companies thrive and which do not.