Unlocking contract insights with natural language processing
Both insureds and policyholders rely on insurance policy and coverage clarity. Artificial intelligence can help.
Large commercial and specialty insurance is a complex, document-intensive business. The complexity of policy wordings — the contract coverage, endorsements, exclusions, terms and conditions embedded within contracts — has grown exponentially.
Achieving contract certainty, clarity on policy terms and conditions that reduces or eliminates coverage ambiguity, is critical to the solvency of both insureds and policyholders. Lloyd’s of London estimates that disputed claims, including legal fees and unplanned claims payments as a result of inaccurate wordings, cost the market about $600 million annually.
Highly experienced underwriters and specialists review contracts, but as contract verbiage becomes more scripted and less standardized, it can be challenging for even a specialist to know where to focus. The National Law Review notes, “It is not uncommon for the basic policy provisions to be subject to exclusions. There can also be exceptions to those exclusions and there may be policy endorsements that override everything. It can be a very difficult maze to navigate.”
Words matter
A shortage of linguistic talent within the insurance industry compounds the problem, with fewer human experts available to determine the content and intent of policy wordings. A 2017 Lloyd’s survey discovered that only 50% of managing agents perform a thorough review of 100% of contract wordings prior to bind, and 40% have no dedicated wordings experts on hand.
These resources are key in helping to determine if wordings are understandable, of good quality, and alignment with underwriting guidelines.
Additionally, most insurers store issued contracts within electronic document management systems, but the detailed information in those documents is not easily accessible. Any portfolio-level review of coverage to identify exposure or trends across requires a manual, time-consuming review of contracts.
Freeing words from the page
Recently, artificial intelligence solutions using natural language processing (NLP) have emerged to assist with the evaluation of contract policy language. Capabilities like intelligent robotic process automation (RPA) and NLP have long been used to automate the extraction of information from standardized documentation.
However, within the large commercial and specialty industry, lack of standardization makes it more challenging to apply NLP capabilities.
At AXA XL, we are piloting an NLP wordings evaluation initiative. We start with a series of business questions that we want to ask and identify documents likely to contain that information. We upload and scan the documents using Optical Character Recognition (OCR) processes to convert text into a computer-readable format.
At this stage, the computer doesn’t have context on what the extracted data points are or their relevance. We start applying NLP or machine learning techniques to help contextualize data elements, but human intervention is still needed to assign meaning.
Once humans “train” the system by validating or providing context, the computer intelligently learns and can start to appropriately categorize both documents and text elements. We can now query contracts in human language (a process called natural language understanding, or NLU), like how you might ask a question in Google.
A user types in “Is flood covered?” and a list of policies with flood coverage will appear. The user continues asking questions and filtering on different taxonomies within the document, such as specific limits or sub-limits associated with the coverage. The ability to automate this process allows us to increase the breadth of documentation for review and increase both speed and accuracy in finding specific answers.
What the future looks like
There is no limit to the number of opportunities to develop our NLP wordings evaluation capability. Here are a few key areas of interest:
- Identifying coverage change across the portfolio: At the account level, NLP facilitates the process of comparing policies to monitor changes in coverage language over time. At the portfolio level, we can uncover coverage trends across different business segments, industries or occupancies, to name a few. This provides us with insights and opportunities to develop new insurance product and coverage solutions.
- Finding ‘Black Swans’: A “black swan” is an unpredictable event that has catastrophic impact. From a risk management perspective, we want to be able to perform worst-case scenario tests for black swan events. This capability would allow us to quickly pull contracts containing coverage for the hypothetical event. While we do this work today, we have an opportunity to increase the breadth and depth of the scenarios that we want to evaluate.
- Post-event evaluation: Underwriting managers can identify coverage issues in their portfolio post-event. As part of a claims lessons-learned process, underwriting may want to see where they have accounts with a similar clause / wording that resulted in the loss. This process helps us determine good wording standards for future policies. Additionally, claims adjusters will be able to quickly access coverage areas in policies to determine if the policy covers the loss.
NLP provides a critical capability for uncovering insights in our document-intensive business. By utilizing these tools, we’re able to capture and analyze documentation, which frees up our experts to ask more questions and find answers quickly.
Rachel Alt-Simmons (rachel.altsimmons@axaxl.com) is head of Business Architecture and Design at AXA XL. Her co-authors are Steven Walden, director, Strategic Operations, Global Property at AXA XL; Bob Lavoie, senior product manager, XL Catlin; and James Breeze, Digital Artificial Intelligence & Analytics lead, XL Catlin.
Other articles by Rachel Alt-Simmons: