Building OCS’s Multi-Modal Human in the Loop Claims Agent

About OCS

One Claim Solution (OCS) helps restoration contractors get paid faster. Founded in 2016, OCS addresses the issue of contractors waiting months for payment after disaster-related rebuilds. Since 2016, OCS has processed over 50,000 invoices, comparing contractor submissions against insurer estimates, identifying discrepancies, and drafting justifications that help contractors get paid faster. 

But as OCS grew, the process that made them successful became their biggest bottleneck. With the help of Focused, OCS clients now average payment in 32 days, significantly faster than industry standards.

The Challenge: Manual Work That Didn’t Scale

Invoices arrived as PDFs, images, and scanned documents with inconsistent structures, making automation unreliable. Existing tools lacked the accuracy and nuanced reasoning of human experts. Focused partnered with OCS to develop a system that handled documentation volume and accuracy, empowering, not replacing, their human experts.

As CEO, Dan Doud put it:

We didn’t want our people spending time comparing PDFs. We wanted them focusing on justifying differences. Focused gave us that.”

The Solution: A Multi-Modal AI Agent Working alongside Claims Advocates

The solution was a multi-modal, human-in-the-loop agent that augmented advocates rather than replacing them.

Focused rapidly built an Ambient Agent using LangGraph, integrating LLMs and AI into OCS's environment with these advanced features:

Tool Calling

OCS runs its business on top of a home-built claims management system that exposes APIs for other services. In order for the AI Agent to work alongside claims advocates Focused built custom tools that enabled the Agent to read and write into OCS’ core framework.

Human-in-the-loop

OCS required their agents to have complete control over the AI’s actions. By leveraging HITL each step of the AI Agent’s actions could be approved, edited, or rejected by a human claims advocate.

​​Replay and resubmission

Each individual step the agent takes matters, By combining HITL and checkpointing claims advocates could easily update, correct, or enhance decisions the AI made and then replay the autonomous actions saving hours of work.

Long-term memory

Every time a human claims advocate guided the agent the AI would analyze what the human did and if the mistake could be avoided in the future. If so, the agent would add a fix to long-term memory allowing the agent to improve over time and in between sessions.

MultiModal subgraphs

Processing complex PDF invoices that included photos, charts, line items and paragraphs of text was core to this agent. By separating the agent into nodes that each serve a purpose the team was able to use Mistral-OCR for document recognition, while leveraging other purpose driven models like ChatGPT and Gemini for reasoning and tool calling.

The LangSmith Difference

Integrating LangSmith from the start provided robust observability (o11y), accelerating development and ensuring high-quality outputs. Human-in-the-loop evals proved more effective than automated LLM judges, ensuring accurate code references and carrier-appropriate language.

During development, LangSmith's debugging and tracing capabilities allowed for rapid iteration and transparent feedback, enabling the creation of evaluators to measure AI output quality and track accuracy.

In production, LangSmith's o11y offered continuous monitoring, performance insights, and proactive optimization, ensuring the reliability and improvement of OCS's AI-driven claims processes.

The Agent

The agent is structured into four key stages: Extraction, Parsing, Matching, and Justifications.

Extraction: Mistral-OCR achieved nearly 90% accuracy in structured data extraction from diverse submission packets (PDFs, images, mixed media) using PyMuPDF for structured parsing and layout preservation..

Parsing: Human-in-the-loop validation allowed advocates to correct extracted fields in real-time, maintaining high accuracy and building trust.

Matching: AI, with human oversight, matched contractor line items to insurer estimates. The system learned recurring terminology differences and improved over time.

Justifications: Automated justification was achieved by grounding the LLM in OCS’s internal knowledge base and industry standards using semantic RAG, supporting advocate reimbursement recommendations.

Enablement, Not Dependance

Focused pair programmed with OCS to build the AI agent and enable OCS developers to build foundational AI architecture to scale future AI efforts. OCS’s engineering team was embedded throughout the process, learning how the system worked and how to extend it and how to maintain it.

CTO Eric Terry added:

Building the foundation with Focused and getting the internal education is a big win for us. Our devs now understand how to extend what we built together.”

This approach gave OCS not just an AI tool but an AI platform they could own and evolve.

Real Results, Real Adoption

OCS claims advocates now daily use a tool that consolidates data and pre-generates justifications, replacing hours of document reconciliation. This foundational agentic system transformed OCS, providing the tools and architecture for future AI workflows.

Doud summed it up best:

The tool and the other stuff we intend to build with AI will eventually allow us to drive down costs to serve, which will let us price at a point that’s almost impossible to match. We got a great kickstart down that path with Focused.”

The Focused Difference

We don’t build prototypes. We build production systems.

At Focused, our agents combine reasoning, retrieval, and real execution. They fit into existing workflows, handle real data, and scale as teams grow.

The OCS project is a clear example. Together, we built a domain-specific system that improved accuracy, reduced manual effort, and created a foundation for future agents.

We build systems that think with you, learn from your work, and make real progress possible.