AI Architecture for Intelligent Document Processing (Data Extraction) in Underwriting and Claims Handling

What inspired me to write this article were discussions with our clients and my team members, who were often focused on using Large Language Models (LLMs) for almost every use case, in contrast to classical Deep Learning (DL) models. LLMs are powerful and easy to use, but it is worth examining their limitations and understanding where it makes sense to apply them — or not — within an overall AI architecture.

This article describes an approach to processes such as underwriting or claims handling, where extracting structured data — an essential part of Intelligent Document Processing (IDP) — is key to automation. It explores the business perspective on touchless processing and Human-in-the-Loop (HITL), and how these requirements influence the selection of AI models and, more broadly, the design of AI architectures.

Moreover, AI tools are sometimes broadly and incorrectly associated with an AI-agentic approach. This article therefore also looks at the relationship between the two and discusses when an agentic approach is actually needed in the context of IDP.

First, let’s clarify why automating document data extraction matters at all in the insurance context. Here, by automation we mean the full spectrum - from fully touchless processing to strong human augmentation.

The Sollers AI Report, released in 2025 and based on in-depth interviews with insurers, shows that IDP is one of the most important trends in AI-driven automation in insurance. However, at present, there is more progress in claims handling than in underwriting.

Beyond the report, at Sollers we have developed a vision for end-to-end automation of the underwriting process. In this vision, smart submission — currently the main focus across the market — is only the starting point.

The worst-case HITL scenario involves asking a user to verify numerous data fields extracted by AI, while providing only a long PDF document with no guidance on where to look. In this case, the user must spend significant time searching and verifying, and their attention becomes fragmented — ultimately reducing verification accuracy.

Whether HITL is used as an exception-handling mechanism for touchless automation or as a standard workflow where every AI recognition is reviewed to increase productivity and/or quality, the user requirements for effective verification of AI output are largely the same.

Users should be given a clear, high-level overview of the case status and know exactly what actions are required from them. They should not be expected to spend time verifying unnecessary information. Instead, their attention should be focused on validating data fields that are critical to business outcomes, such as risk assessment, policy pricing, or claim payouts.

Classical Deep Learning models are neural networks trained for a specific, well-defined task, such as image classification, speech recognition, or document data extraction. They typically learn from labelled datasets, have fixed inputs and outputs, and, once trained, their behaviour does not change unless the model is retrained..

A Large Language Model is a single, very large Deep Learning model — typically based on the transformer architecture — composed of many stacked neural network layers and billions of parameters. LLMs are trained on massive volumes of text to learn the general structure and patterns of language.

Unlike classical DL models that are optimized for a single task, LLMs can generate, understand, summarize, and reason over text across many tasks using the same underlying model. They can be adapted to new tasks through prompting or fine-tuning, rather than full retraining.

For high-volume, well-defined document extraction tasks, task-specific DL models generally require less compute per inference than LLMs. As a result, they can be more cost-effective when deployed at scale.

However, factors such as licensing, maintenance, and flexibility may make LLM-based solutions more economical for low-volume or highly variable use cases.

Mature DL models typically provide out-of-the-box confidence level (CL) metrics for each extracted data field and label. These metrics enable the use of statistical methods to filter out uncertain recognitions and route them to manual (human) handling.

LLMs, in contrast, do not provide such confidence levels out of the box. While it is possible to ask an LLM to estimate confidence, the general-purpose nature of these models means such values often do not reliably fulfill the role of true confidence metrics. Other techniques exist to control recognition correctness, but they (a) are often not fully effective and (b) require additional effort to develop and maintain.

A similar comparison applies to source traceability. Mature DL models usually provide out-of-the-box references to the source document, including coordinates of the extracted data. This enables fast and accurate verification.

LLMs do not inherently provide this information. Although various techniques can be used to approximate source references, they are not always reliable and typically require additional implementation and maintenance effort.

When starting a Proof of Concept (PoC), LLMs are usually the simplest choice. They are easy to use, allow teams to see initial results quickly, and often deliver very promising outcomes. However, as shown so far, when a more holistic view of the target process is taken, the picture becomes more complex.

The diagram illustrates various factors indicating which type of AI—DL or LLM—tends to be more suitable for a given scenario. Ultimately, the final decision should be driven by a holistic business case. At the beginning, an organization may not yet have sufficient knowledge or data to fully inform that business case. This understanding develops over time through practical experience — by experimenting, testing, implementing, and inevitably learning from mistakes.

Extracting data from documents is only one part of the overall IDP process. Although it is the most common focus for PoCs, a target end-to-end process typically requires additional activities that can also be automated using AI, such as document separation, deduplication, classification, data extraction, cleaning, and validation.

To provide a complete perspective, the comparison must also include business rules (or business logic in a broader sense), which play a critical role alongside DL models and LLMs. To give an initial overview, the diagram presents typical examples of how business rules, DLs, and LLMs can be applied across the full IDP lifecycle.

Each of these components plays a different role and has a different level of importance depending on the task. Business rules ensure determinism, control, and compliance; classical DL models deliver scalable and accurate perception and extraction; and LLMs provide semantic understanding and adaptability in ambiguous scenarios. The most mature IDP platforms are hybrid: together, these components enable high levels of automation while maintaining reliability and governance in enterprise IDP processes.

Effective implementation requires solid tools, but making the right architectural choices for AI-driven automation requires experience-based knowledge of what it truly takes to automate processes with AI. The challenge is that this knowledge is typically gained only through hands-on implementation.

Engaging experienced experts or consultancies can accelerate implementation by bringing in proven knowledge and best practices. However, it is equally important to use this opportunity to help internal implementation teams learn and mature quickly. Insurers ultimately need this internal expertise to make sound architectural and transformation decisions.

For organizations that are new to AI, designing a mature, company-wide transformation from the outset is extremely difficult without prior experience.

Begin by building a high-level vision of the required target architecture and its components. Start implementation with simpler tools (for example, cloud-provider AI services), and plan to make more informed decisions about advanced components — such as IDP platforms or AI-agentic platforms — once the organization has gained sufficient practical experience.

With a long-term vision in place, adopt a phased approach in which continuous learning is combined with gradual development of architecture and AI-driven business logic. Insights gained from implementation projects and strategic planning should reinforce each other — helping to identify areas with the greatest automation potential and to introduce increasingly advanced toolsets, phase by phase.