INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026

Agentic AI: Insurance Claim Processing System

delays of 24–48 hours per claim. This paper presents an Agentic AI Insurance Claim Processing System that

utilizes a collaborative multi-agent architecture to automate the end-to-end verification pipeline. Powered by the

LLaMA 3.1 (8B) large language model via the Groq API, the system orchestrates four specialized agents—

Policy, Fraud, Eligibility, and Decision Agents—to validate claims against structured Excel datasets.

Experimental results on a curated test dataset of 30 insurance claims demonstrate a 93.33% accuracy rate with a

mean processing time of 78 seconds, well within the 2-minute operational threshold, demonstrating the viability

of Agentic AI in real-world financial services automation. The per-agent reasoning logs were independently

assessed by insurance domain experts and confirmed to meet regulatory audit sufficiency standards, directly

addressing the critical explainability gap identified in prior literature.

Keywords: Agentic AI, insurance claim processing, multi-agent system, LLaMA 3.1, fraud detection,

multiple disconnected systems, resulting in operational inconsistencies, high administrative costs, and frequent

delays of 24–48 hours per claim.

Despite the emergence of robotic process automation (RPA) and rule-based decision engines in recent years,

most deployed solutions remain fragmented. They address isolated sub-tasks within the claim lifecycle rather

than providing a comprehensive, end-to-end automation pipeline. Furthermore, the lack of transparent reasoning

in existing automated systems creates significant compliance and audit challenges for insurers operating in

heavily regulated markets. To address these critical limitations, this research introduces a modular Agentic AI

framework built on a collaborative multi-agent architecture. By deploying specialized intelligent agents that each

own a discrete verification domain—policy validation, fraud analysis, eligibility assessment, and final

adjudication—the system transforms the traditionally opaque and error-prone claim processing workflow into a

and fraud detection. Their study assessed twenty-three distinct AI deployment architectures across major

insurance providers and concluded that while ML models significantly outperform traditional rule-based systems

in predictive accuracy, they consistently fail to provide the degree of decision transparency required for

regulatory compliance. The authors identified this explainability gap as the most critical barrier to widespread

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026

AI adoption in insurance adjudication. This limitation directly motivates our design of a multi-agent system

wherein each agent generates a detailed, auditable reasoning log for every decision made during the claim

verification lifecycle, ensuring full traceability from submission to final adjudication.

Wu et al. (2023) presented an empirical study of Agentic AI workflows applied to financial decision-making

tasks, analyzing the performance of multi-agent decomposition strategies across equity trading, credit

assessment, and insurance underwriting domains. Their research conclusively demonstrated that decomposing

Rawat (2024) explored the use of Large Language Models for automating insurance claim validation using

structured JSON datasets, demonstrating that LLMs can effectively parse structured tabular data, validate multi-

condition claim eligibility rules, and generate natural-language justifications for adjudication decisions without

the need for custom-trained domain-specific models. However, the proposed single-model architecture lacked

agent specialization, resulting in higher rates of cross-domain reasoning errors particularly in fraud detection

scenarios. Our system directly extends this work by distributing claim validation responsibilities across four

dedicated agents, each handling a distinct, bounded reasoning domain with specialized prompting strategies.

Brown et al. (2020) demonstrated that large-scale language models exhibit strong few-shot learning capabilities,

enabling effective task completion with minimal domain-specific fine-tuning. This foundational finding

validates our use of the LLaMA 3.1 (8B) model for multi-domain claim verification without requiring custom

informed our design decision to leverage LLM-based reasoning in the Fraud Agent, enabling dynamic, context-

aware fraud assessment that can generalize beyond the fixed pattern distributions captured by static ML

classifiers.

Open Issues and Research Challenges

A. Data Quality and Completeness

Insurance claim datasets are frequently incomplete, inconsistently formatted, or contain erroneous entries arising

from manual data entry errors, legacy system migrations, and cross-departmental data silos. Agents operating

on such data may produce inaccurate policy validations, missed fraud signals, or incorrect eligibility

determinations. Developing robust data preprocessing pipelines and real-time anomaly detection sub-routines

capable of dynamically identifying and gracefully handling missing or malformed claim records remains a

responsive to novel, previously unseen fraud indicators—without requiring full retraining of the underlying

LLM—necessitates the development of adaptive, knowledge-augmented reasoning strategies that can

incorporate emerging fraud intelligence in near-real time.

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026

C. LLM Hallucination and Output Determinism

Large Language Models are inherently probabilistic generative systems, meaning that semantically identical

inputs may occasionally produce subtly different outputs across multiple inference calls. In the high-stakes

domain of insurance claim adjudication—where decisions carry direct financial and legal consequences—this

non-determinism poses a significant operational risk. Implementing strict output schema validation layers,

structured response enforcement, and deterministic post-processing pipelines on top of raw LLM outputs is

essential for ensuring the consistency and legal defensibility of agent decisions across all claim processing

orchestration architectures with dynamic horizontal scaling capabilities represents a critical engineering

challenge for production-scale enterprise deployments.

E. Regulatory Compliance and Explainability

Insurance claim adjudication operates within a complex and jurisdiction-specific regulatory environment that

mandates transparent, documented, and auditable decision-making processes. Ensuring that the natural-language

reasoning logs generated by each agent are not only human-readable but also fully compliant with applicable

jurisdictional insurance regulations—including IRDAI guidelines and GDPR data handling requirements—is a

non-trivial challenge that future iterations of the system must address through the integration of regulatory policy

parsing and compliance cross-referencing modules.

F. Integration with Legacy Insurance Systems

Proposed Solution

The system is organized into four tightly integrated functional agents, each responsible for a discrete, non-

overlapping verification domain within the insurance claim lifecycle. The agent pipeline mirrors the sequential

decision-making process employed by human claim verification teams, decomposing the overall adjudication

task into individually auditable, domain-specific reasoning steps.

A. Policy Agent

The Policy Agent serves as the first verification gate in the claim processing pipeline. Upon receiving a submitted

claim, the agent queries the structured Policy Excel dataset to verify the existence of the policyholder record,

validate the current active or inactive status of the policy, and confirm whether the requested treatment type is

explicitly covered under the policyholder's subscribed coverage tier.

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026

B. Fraud Agent

The Fraud Agent performs a multi-dimensional fraud risk assessment by analyzing the claim record against

historical submission patterns stored in the Claims dataset and evaluating the completeness of supporting

documentation recorded in the Documents dataset. The agent applies a scoring heuristic that weights anomaly

indicators including duplicate claim submission patterns, claim amounts that materially exceed the statistical

distribution for the associated policy tier and treatment category, and missing or incomplete document

submissions, assigning a fraud risk classification of Low, Medium, or High. The fraud determination is derived

reasonableness against the applicable coverage limit. All six criteria must be simultaneously satisfied for a claim

to advance to the Decision Agent. The agent reports each criterion as PASS or FAIL, ensuring full transparency

D. Decision Agent

The Decision Agent represents the terminal stage of the pipeline, synthesizing the structured outputs of all three

preceding agents to generate the definitive claim adjudication determination. For approved claims, the agent

confirms the disbursement amount equal to the claimed amount subject to applicable coverage limits. For

rejected claims, the agent generates a comprehensive rejection notification that clearly articulates each specific

validation criterion that was not satisfied, along with actionable guidance for resubmission.

System Architecture

Data Layer: The system operates directly against a structured Excel workbook serving as the authoritative data

source. The workbook contains three discrete sheets: the Policy Sheet storing policyholder identity records,

subscription tier classifications, and treatment coverage mappings; the Claims Sheet maintaining submitted

claim records including treatment codes, claimed amounts, fraud flags, and historical claim counts; and the

Documents Sheet tracking the submission status of all required supporting documentation for each active claim.

Agent Orchestration Layer: The four agents are implemented as independent Python modules, each

encapsulating its own domain-specific system prompt, input data retrieval logic, and structured output parsing

routine. The orchestration layer manages sequential agent invocation, passing the structured output of each

completed agent as context to the subsequent agent, enabling coherent multi-step reasoning across the full claim

lifecycle.

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026

LLM Hallucination: As identified in the error analysis, the LLaMA 3.1 8B model occasionally misreads

treatment coverage lists, producing false positive approvals. Mitigation strategies include implementing strict

output schema validation and deterministic post-processing layers. Static Fraud Detection: The current fraud

detection relies on pre-computed dataset flags rather than dynamically trained anomaly detection models,

limiting adaptability to evolving fraud patterns. Integration of graph neural network-based fraud analytics is

identified as a priority for future development. Data Privacy: The current system transmits claim data to the

Groq external API for inference, raising potential data privacy and regulatory compliance concerns in production

environments. Future iterations will evaluate on-premise LLM deployment options to address this limitation.

This paper has presented the Agentic AI Insurance Claim Processing System, a novel multi-agent framework

94.7%, Recall: 94.7%, F1: 94.7%) with a mean end-to-end processing time of 78 seconds, representing a

substantial improvement over the 24–48 hour manual processing benchmark. The confusion matrix analysis

confirmed 18 true positives, 10 true negatives, 1 false positive, and 1 false negative. The per-agent reasoning

logs generated by the system were independently assessed by insurance domain experts as meeting regulatory

audit sufficiency standards, directly addressing the critical explainability gap that has historically impeded

enterprise AI adoption in the insurance domain.

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026

Future development directions include: (1) expansion of the experimental evaluation to larger and more diverse

insurance datasets; (2) integration of graph neural network-based fraud pattern analysis to replace static heuristic

fraud scoring; (3) implementation of privacy-preserving AI strategies and on-premise LLM deployment to

address data security concerns; (4) deployment of asynchronous multi-agent orchestration for scalability under

concurrent claim loads; (5) formal benchmarking against commercial claim automation systems; and (6)

integration of adaptive reinforcement learning mechanisms for dynamic eligibility threshold calibration.

REFERENCES

4. T. Brown et al., "Language Models are Few-Shot Learners," Advances in Neural Information Processing

Systems (NeurIPS), vol. 33, pp. 1877–1901, 2020.

5. A. Vaswani et al., "Attention Is All You Need," Advances in Neural Information Processing Systems

(NeurIPS), vol. 30, pp. 5998–6008, 2017.

6. R. Zhao and K. Liu, "Fraud Detection in Insurance Claims Using Ensemble Machine Learning Models,"

Expert Systems with Applications, vol. 195, pp. 116–128, 2022.

7. P. Kasneci et al., "ChatGPT for Good? On Opportunities and Challenges of Large Language Models for

Education," Learning and Individual Differences, vol. 103, 2023.