Building on the architectural foundation of agentic AI and deterministic workflows from my previous post, I have developed a functional demo tailored to a high-impact industry use case: Insurance Claims Processing.
While this prototype is not a “full-scale” production environment, it is designed to tackle some of the critical bottlenecks in customer service and demonstrate the tangible potential of autonomous agents available 24×7.
Verified autonomy: Scaling expert impact
The primary goal of this solution is resolving routine inquiries and conducting initial assessments before a human expert is brought into the loop. All this to reduce the backlog, and release capacity by offloading tasks from claims adjusters.This solution isn’t just about answering questions; it’s about end-to-end task handling for existing policyholders. By allowing AI to autonomously handle routine, “straightforward” inquiries, the system creates the necessary bandwidth for human experts to focus on what they do best: the nuanced evaluation of complex claims and critical validation steps that demand human oversight.
The demo environment covers a comprehensive suite of typical customer touchpoints and workflows, including:
- Data management: Querying or updating personal policyholder information
- Policy intelligence: Retrieving details on existing coverage and terms
- Claim lifecycle: Intake of new claims, automated classification of claim types, and status tracking
- Coverage verification: Cross-referencing submitted claims against active policies. The policy types range from Health, Car, Home, Life, Renters, to Disability
Beyond simple chat: Context and multimodality
Going further, the agentic system can request and process specific information—such as photos of physical damage, estimated repair costs, or missing information required for a new policy. A key feature is contextual continuity: customers can simply reply to previous email threads, and the agent will utilize the entire history as relevant context for the next step.
Empowering the expert: The client-relationship dashboard
Automation does not mean a lack of oversight. I have implemented a centralized dashboard that allows claims adjusters to maintain full visibility. The system intelligently flags tasks that require human intervention—such as the final valuation of a loss, payment approvals, or the appointment of a specialized appraiser. This ensures that complex cases and high-stakes decisions are seamlessly “triaged” to a human expert (Human-in-the-Loop (HITL)).
Real-world scenarios in action
Seeing is believing. To demonstrate how the agentic system handles the complexities of insurance customer service, I have curated two demo scenarios. These videos showcase an end-to-end journey—from the initial customer email to the final resolution or handover to a human expert.
Handling a smartphone claim
In this fully autonomous demonstration, we follow a customer reporting physical damage to their smartphone. Notably, the customer provides no written description of the incident; instead, they simply submit two photos of the damaged device as attachments.
The agentic system performs a multimodal analysis of the images to identify the nature and extent of the damage. Once the context is established, the agent cross-references the findings against the customer’s existing insurance policies.
Upon review, the agent determines that there is no policy which covers accidental hardware damage. However, rather than delivering a simple rejection, the system utilizes its deterministic reasoning to turn a service gap into a value-add. The agent proactively informs the customer that the current claim cannot be covered but immediately presents a tailored proposal for a new supplemental policy, including a clear breakdown of terms and premiums to cover such incidents.
The customer accepts the offer to ensure they are fully protected against future damage. Without any human intervention, the agentic workflow executes the entire backend process: it generates the new policy, integrates it into the customer’s active portfolio, and triggers a formal confirmation. On the administrative side, the insurance adjuster can see the updated coverage and the successful transaction reflected in the client-relationship dashboard.
Handling a car damage claim
In this scenario, an existing client initiates the process by emailing the service desk with their client ID and a photo of their vehicle’s damage. Upon receipt, the system performs an automated image analysis to translate the visual data into a detailed textual description.
For example, the AI might generate the following assessment: “The attached image indicates a moderate-to-high severity side-impact collision on the rear passenger-side door, characterized by deep scratches, surface rust, and minor panel deformation.”
The system automatically cross-references the damage description against the client’s active insurance policies. In this case, the AI determines that the damage is covered under the client’s car policy. An automated email is sent to the client confirming coverage and inviting them to open a new claim.
The client responds directly to the email, confirming their request and attaching a repair shop invoice. The system utilizes Intelligent Document Processing (IDP) to extract and validate key line items from the invoice, specifically identifying the total amount of 2541,40 Euro. Once the data is extracted, the client receives a comprehensive summary via email. This notification includes:
- Claim ID and the linked policy
- Claim type and the validated amount of 2541,40 Euro
- Description: Merging the initial image analysis with the line-item details found in the repair shop invoice
- Timeline: A note stating that the claim will undergo a manual review, with a final decision expected within the next two business days
On the backend, the claims adjusters accesses the client-relationship dashboard to review the newly generated claim. The file is automatically flagged with the status “human-review”, allowing the employee to easily pick it for further processing and approval.
How it works – Journey of every customer interaction
While the results may feel like AI “magic”, the system is actually grounded in a robust and logical data pipeline. The core strength of this solution lies in its ability to bridge the gap between messy, unstructured customer input into structured, actionable data points required for automated service delivery and claims processing.
1. Ingestion and extraction
The process is triggered when a customer sends an email to the insurance helpdesk. A specialized email poller acts as the gateway, instantly identifying new messages.
The poller extracts the subject line, body text and all attachments. The system accepts documents in various formats, including PDF, DOCX/DOC, PPTX/PPT, Excel, and CSV. From these files, it identifies and extracts critical data points such as:
- Customer, policy, and claim IDs
- Contact information (email addresses and phone numbers)
- Dates and monetary amounts
For attached images (JPG, PNG, HEIC, GIF, BMP, TIFF), the system leverages Google Gemini 2.5 Flash for automated damage assessment. It generates detailed descriptions of visible damage, creates severity ratings, and provides initial repair scope estimates. Finally, all information from the email, attachments, and image analysis is compiled into a single, context-enriched prompt.
2. Analysis
This enriched context is handed off to IBM watsonx Orchestrate. Serving as the “agentic core,” it uses the extracted information to automatically classify the request and query the Customer Relationship Management (CRM) system for client information, relevant policy details and claim histories.
It then applies deterministic logic and guidelines to decide the next best action—whether that is creating a new claim, offering a new policy, retrieving specific client information or escalating to a human expert. Under the hood, Orchestrate uses GPT-OSS 120B (via Groq) as its primary language model, using integrated tools to list, create, or update information directly within the CRM.
3. Response
The final step of the interaction is the generation of a professional, personalized response that includes the relevant email history.
This answer is sent back to the original sender via the email poller. To ensure a seamless experience, the poller maintains a conversation thread, allowing the system to accurately map the interaction to the specific customer and respond to any follow-up inquiries with the full context of the history.
Conclusion
The agentic system handles routine inquiries and initial assessments to cut manual overhead, while maintaining a human-in-the-loop for all critical validations and decisions. As mentioned initially, this system is a prototype focusing on specific workflows. It can be further customized and expanded as needed.
In this context, there is a constant trade-off regarding the degree of autonomy to be granted. Building trust in such solutions is essential, and the most effective way to achieve this is by first automating routine, low-risk workflows.
As an article by EY points out, agentic AI in claims management can serve as a “… lever for end-to-end automation and genuine leaps in efficiency.” The authors highlight—among other things—automated initial assessments, the processing of standard cases, and document and image analysis as key efficiency drivers, all of which have been practically implemented in the prototype presented here.