Insurance is inherently document-intensive:

Every day, carriers and MGAs process large volumes of applications, ACORD forms, Loss Runs, Statements of Values (SOVs), supplemental applications, policy documents, endorsements, claims reports, estimates, invoices, inspection reports, correspondence, spreadsheets and other supporting documents.

For years, OCR and Intelligent Document Processing (IDP) have helped organizations digitize documents and reduce manual data entry. However, the challenge facing insurers is evolving.

The question is no longer simply, “Can we extract information from this document?” Increasingly, insurers are asking, “Can we understand what these documents are telling us and help our insurance professionals act on that information?”

This represents an important shift from document processing to Document Intelligence. It is to create an adaptable intelligence layer that can understand different documents, apply business context, identify what matters and deliver that intelligence into the workflows where insurance professionals make decisions, for achieving this intelligence requires more than a standardized, one-size-fits-all approach.

Why Insurance Documents Are Different

Consider something as familiar as a Loss Run.

The business information may appear straightforward:

• Claim number
• Loss date
• Cause of loss
• Paid amount
• Outstanding reserve
• Total incurred
• Claim status

Yet the way that information is represented can vary significantly.

Different carriers may use different layouts, terminology, table structures, page structures and reporting conventions. One report may present all information in a structured table, while another may distribute important information across multiple pages and narratives.

The challenge becomes even greater when processing documents such as SOVs, commercial submissions, claims files, inspection reports or supplemental applications. Documents may contain:

• Structured information in clearly defined fields.
• Semi-structured information in tables and schedules.
• Unstructured information within descriptions, notes, emails and narratives.
• Information required for a business decision that exists across several documents rather than inside one document.

This is why simply selecting a document-processing product and configuring a few fields may not address every insurance use case.

The Challenge Is Bigger Than Extraction

A traditional document-processing requirement might be to “extract these 25 fields from this document.” While that remains useful, insurance organizations increasingly have more sophisticated questions:

• Underwriter: “What are the major losses during the last five years?”
• Property underwriter: “Are the location values in the SOV consistent with the application?”
• Renewal underwriter: “Tell me what materially changed from last year's submission.”
• Claims examiner: “What changed in this claim during the last 60 days?”
• Operations team: “Is this submission complete enough to move to underwriting?”

These are no longer purely document-extraction problems. They require the ability to understand, compare, validate, summarize and analyze information in an insurance context.

Why a Customized Approach Matters

Every carrier and MGA has developed its own combination of products, underwriting guidelines, operating processes, data requirements and technology platforms. As a result, the intelligence required from documents also differs.

• One carrier may need to process thousands of commercial submissions and determine whether required information has been received.
• Another may want to normalize Loss Runs received from a wide variety of carriers.
• An MGA may want to accelerate submission-to-quote for a specific specialty program.
• A commercial property insurer may be focused on validating SOV information.
• A claims organization may need to help examiners navigate hundreds of pages of claim documentation.

The underlying AI capabilities may be similar, but the business solution is different. A practical document-intelligence solution therefore needs to adapt across several dimensions:

• Document Types: What documents need to be understood?
• Data Requirements: What information needs to be extracted and normalized?
• Business Rules: What validations, calculations and exceptions need to be applied?
• Intelligence Requirements: What comparisons, insights, summaries or questions should the system support?
• Quality Requirements: What confidence levels require validation or human review?
• Operational Requirements: What volumes and processing SLAs need to be supported?
• Integration Requirements: Where should the resulting information and intelligence be delivered?

This is where customization becomes important.

From Document Processing to an AI Intelligence Layer

At Kumaran Systems, our work with complex insurance documents has helped us understand that document processing is not simply about recognizing text. It requires progressively understanding the structure, context and business meaning of the information being processed. That experience is shaping our approach to Kumaran Insurance Document Intelligence (KIDI).

Rather than treating KIDI as another fixed document-processing product, we see it as an AI Intelligence Layer that can be configured around an insurer's or MGA's document-processing use cases.

The evolution of the KIDI AI Intelligence Layer can be viewed in three stages:

1. Document Intelligence

Classify → Extract → Normalize → Validate

The first stage reliably transforms different document types and formats into usable insurance data.

2. Business Intelligence

Compare → Detect → Summarize → Analyze

Once information is structured, intelligence can be applied to:

• Identify inconsistencies between documents.
• Detect significant changes between renewal periods.
• Analyze loss frequency and severity patterns.
• Flag missing information, unusual exposure changes, and data requiring further review.

Consider an underwriter reviewing five years of Loss Runs.

Instead of simply receiving extracted claim records, the underwriter could receive an insight such as:

Loss activity increased during the most recent two years, with three locations accounting for a significant portion of total incurred losses and water damage representing the most frequent cause of loss.

The information has not simply been extracted, it has been transformed into business intelligence that helps direct the underwriter's attention.

3. Conversational Intelligence

Ask → Understand → Investigate → Act

This layer allows insurance professionals to interact naturally with the information. Instead of manually searching through a 100-page submission, an underwriter could progressively ask:

• “What are the most significant losses?”
• “Are those losses concentrated at particular locations?”
• “What information does the insured provide about corrective actions?”

A claims examiner might ask:

• “What changed in this claim during the last 60 days?”
• “Why did the repair estimate increase?”
• “Show me the supporting document.”

This creates a Co-Assistant model in which AI helps insurance professionals navigate information while the human remains responsible for the business decision. Importantly, “Act” does not necessarily mean autonomous underwriting or claims decisions; it can involve preparing broker questions, generating summaries, creating exceptions, routing information for review, or passing validated information into an existing workflow.

Built from Understanding Document Complexity

An AI intelligence layer becomes more useful when it is built on practical experience with document variation. Processing different formats reveals challenges that may not be obvious in a controlled demonstration, including:

• Tables continuing across multiple pages and changing column structures.
• Inconsistent terminology, missing values, and poor-quality scans.
• Multiple entities within the same document and values appearing in different sections.
• Narratives that provide context to structured fields.
• Conflicting information between documents.
• Document packages containing multiple document types.

Solving these challenges creates reusable capabilities for document classification, extraction, normalization, validation, exception handling, and contextual understanding. Kumaran's objective is to use this experience as a foundation that can be configured for additional carrier and MGA use cases, rather than starting from scratch for every new document-processing problem.

What Could This Look Like for a Carrier?

Consider a commercial carrier receiving a broker submission containing an email, ACORD application, Loss Runs, SOV, and Supplemental Application. The intelligence layer can be configured to:

• 1. Understand the submission: Identify and classify the documents received.
• 2. Extract relevant information: Capture the information required for the carrier's underwriting process.
• 3.Normalize the information: Convert different document formats into a consistent insurance data structure.
• 4.Validate the submission: Apply carrier-specific business rules and identify missing or inconsistent information.
• 5.Compare documents: Identify discrepancies among the application, SOV, Loss Runs, and supporting documents.
• 6.Generate intelligence: Produce a concise submission summary and highlight areas requiring attention.
• 7.Enable interaction: Allow the underwriter to ask questions across the submission.
• 8.Integrate with the workflow: Deliver validated information and insights into the carrier's existing underwriting environment.

The carrier does not necessarily need another isolated application; the intelligence can become part of the process it already uses.

What Could This Look Like for an MGA?

For MGAs, speed and specialization can make this approach particularly valuable. Rather than configuring a generic document-processing solution for a high volume of submissions, the intelligence layer can be aligned with the requirements of a specific insurance program:

• 1.Receive the broker submission.
• 2.Classify the documents.
• 3.Extract program-specific data.
• 4.Validate completeness.
• 5.Generate Loss Run and exposure intelligence.
• 6.Apply configured business-rule checks.
• 7.Prepare an underwriting summary.
• 8.Deliver the output into the existing MGA workflow.

The same capabilities can be configured differently for another program with distinct documents, data requirements, and underwriting processes. This gives an MGA platform operating multiple specialty programs a common document-intelligence foundation while supporting program-specific requirements.

Integration Is as Important as Intelligence

Document processing should not end with a JSON file, spreadsheet, or another dashboard. To create operational value, intelligence must reach the user at the appropriate point in the workflow. Depending on the use case, the output may need to:

• Populate an underwriting workbench.
• Update a policy or claims system.
• Feed a data platform.
• Trigger an exception workflow or create a task for human review.
• Provide an API response to an existing application.
• Present a conversational assistant inside the user's workflow.

This is another reason customization matters: the document may be common, but the workflow rarely is.

Human Expertise Remains Part of the Design

AI intelligence does not mean every insurance decision should become autonomous. Some information may be processed with high confidence, while other information may require validation, human interpretation, or professional judgment. A configurable intelligence layer can support the following flow:

• 1.AI processing
• 2.Confidence assessment and validation
• 3.Business-rule application
• 4.Exception identification
• 5.Human review where required
• 6.Approved downstream action

The objective is to reduce unnecessary manual effort while keeping insurance professionals focused on decisions that require their expertise.

One Intelligence Foundation, Multiple Insurance Use Cases

Once a configurable document-intelligence foundation exists, carriers and MGAs can extend the same core capabilities across multiple processes by adapting them to different documents, business rules, workflows, and decision requirements. The value lies not in building a separate AI solution for every document, but in creating reusable intelligence capabilities that can be configured for different insurance processes.

Your Documents. Your Rules. Your Workflow. Your Intelligence.

The insurance industry is moving beyond simply digitizing documents toward understanding what those documents mean in a business context. This requires a different mindset: rather than starting with “Which document-processing product should we buy?”, organizations can begin by asking “What business intelligence do we want to derive from our documents?”

From there, they can determine:

• Which documents contain that intelligence?
• What processing challenges exist?
• What business rules should be applied?
• What questions should users be able to ask?
• Where should the resulting intelligence appear in the workflow?

Kumaran Insurance Document Intelligence is being developed around this philosophy - leveraging our experience in understanding and processing varied insurance documents to create a configurable AI Intelligence Layer for carriers and MGAs.

The goal is not to force every insurance document into the same predefined process. It is to combine document understanding, insurance context, AI intelligence, and integration capabilities to address the specific business problem surrounding those documents.

Bring Us Your Document Challenge

If your organization is exploring how AI can improve a document-intensive underwriting, claims, or insurance operations process, start with the problem - not the product:

• Show us the documents.
• Tell us the challenges.
• Describe the intelligence your users need.
• Show us where that intelligence needs to fit into your workflow.

Let's explore how a customized document-intelligence solution can help.

Kumaran Insurance Document Intelligence

Different Documents. Different Needs. Intelligence That Adapts