Introduction
With the introduction of tools like AI to extract data from unstructured documents such as claims reports, risk engineering reports, and legal documents, once locked data can now be surfaced, classified and used as a basis for insights and better decision-making.
Claims Data as a Strategic Asset
With the introduction of tools like AI to extract data from unstructured documents such as claims reports, risk engineering reports, and legal documents, once locked data can now be surfaced, classified and used as a basis for insights and better decision-making.
In today’s world of unstructured documents, it is a manual effort to read the documents, key in a handful of pertinent claims information, triage the claim and make coverage decisions. With a focus on causation, what is keyed is often limited to a handful of overly broad categories like “fire” or “flood,” or worse, marked as “unknown” or “other,” meaning the ability to use data meaningfully is not feasible.
Alchemy is working with a client using artificial intelligence (AI) to improve and increase the granularity of the classification of claims causation. From that improved classification, this data can become a strategic asset and unlock valuable insights for multiple use cases such as enhanced policy efficacy, understand claims trends by line of business and targeted risk management and mitigation strategies.
AI is now addressing the challenges of current manual rekeying and classification practices, and in doing so this has positive implications for policy design and risk prevention.
Why is it a problem today?
In addition to the process of managing claims, (portfolio) claims data should inform underwriting, shape policy wording, and support loss mitigation efforts (frequency as well as severity). This is however limited by today’s lack of granular and manually keyed data. Alchemy has observed the common challenges of this limited dataset:
1. Overly Broad Categories
Losses are frequently grouped into generic buckets like “fire,” “flood,” or “cat-related damage.” Whilst these describe the nature of the loss, and are useful to underwriters in their rating models, this does not provide enough detail as to the proximate or underlying cause of loss to enable identification of trends.
2. Generic Words
Classification words like ‘storm’ are also problematic… is this heavy rainfall, a snowstorm, strong winds and so on? Again, this is insufficient detail to make meaningful use of the data.
3. Unknown or Other
Many claims are labelled as “unknown” or “other,” particularly when causation is unclear or when data is missing. Possibly valid for block claims but otherwise, these classification categories provide no actionable information and therefore no ability to understand patterns or trends.
4. Missed Connections Between Causes
Complex losses often involve multiple contributing factors. For example, a workplace injury might involve both inadequate/ poor training and faulty equipment. Current classification systems rarely capture these interdependencies.
It is in understanding the proximate cause of a loss, that insurers can then address the root issues driving claims frequency and severity.
How Can AI Can transform causation classification?
Artificial intelligence is a tool that can extract & analyse structured and unstructured data and provide insights that go far beyond today’s manual process.
1. Granular Classification
AI models can classify claims based on proximate causes rather than generic categories at a far greater speed across a wide range of documents. For example:
- A claim causation of “fire,” is now classified as “overloaded electrical circuits due to inadequate maintenance.”
- For a flood claim, AI could specify “blocked drainage system exacerbated by lack of regular inspections.”
These new granular causations would then need to be classified to an agreed data model for consistency and trend analysis.
2. Pattern Recognition
AI can analyse vast datasets and can then identify recurring patterns and trends. For example, it might detect that a disproportionate number of property claims in a specific region are linked to outdated construction materials.
3. Causation Models
Using advanced causation models, AI can map relationships between multiple contributing factors, providing a holistic view of how and why losses occur.
4. Unstructured Data Analysis
Much of the data relevant to causation is hidden in unstructured formats, such as adjuster notes and reports, legal documents or even risk engineering reports to see if elements had been noted during the risk assessment that subsequently led to a claim. AI can process this information to uncover insights that would otherwise remain inaccessible.
Noting here that these AI tools can look across live claims as well as historic claims. By running AI models and tools over years of historic (and closed) claims that no human could effectively do. This will create a strategic data asset that will be a gold mine for multiple use cases:
The Benefits & Impact of Improved Causation Classification
Improved causation classification has implications for the design and efficacy of coverage and policy wordings. By having accurate data on each risk, this enables:
1. Tailored & Clearer Policy Design & Wording
Granular causation insights enable insurers to create even more personalised and bespoke coverages per clients’ specific exposures. For example:
- A manufacturing company with a history of machinery-related fires could be offered coverage contingent on implementing specific maintenance protocols.
- A property owner in a flood-prone area might receive a policy that includes provisions for improved drainage infrastructure.
Detailed understanding of causation also allows for more precise policy language, reducing ambiguity.
2. Improved Underwriting Decisions
Accurate causation data supports better underwriting decisions, allowing insurers to price risks more effectively and avoid exposures that are outside of their risk appetite.
3. Enabling Risk Prevention
Claims teams that have detailed causation data as well as trend analysis can:
- Provide actionable advice to clients on mitigating specific risks.
- Work with underwriters to ensure that risk prevention measures are factored into pricing and capacity decisions.
- Identify industry-wide trends that inform broader loss prevention initiatives.
For example, if claims data reveals that a significant proportion of workplace injuries in a certain sector are linked to inadequate safety training, insurers could partner with industry bodies to promote training programmes or establish minimum safety standards.
Implementation Challenges of an AI-led Causation & Classification model
Implementation of an AI tool is not without its challenges:
- Data Silos: Claims data and particularly documents are often present in multiple systems and documents stored in multiple locations. It is an effort to consolidate this into one location to run the model against.
- Consistency of Causation Classification: When designing a classification model, it is key to involve multiple human stakeholders, many of whom interpret causation differently in a manual world. Building a consistency of understanding of interpretation within the AI model will improve overall consistency of data output.
- Building Trust in AI Models: Claims teams may be sceptical about relying on AI-generated causation classification, particularly if there is no transparency or traceability of how the models have made their decisions. Insurers must ensure that the software solution and models provide clear explanations of their outputs, and where claims teams validating model outputs (and over time, on a sample basis only). Demonstrating consistent accuracy is key.
Where claims teams can move from generic causation classifications to detailed root causes, this creates data as a strategic asset for more effective policies and proactive risk prevention strategies.