Goldmund's AI Agent Framework

Scott Jones - Front-end Developer & Designer

13 June 2025

At Goldmund, we're developing an internal framework that enables us to deploy AI agents for various tasks and workflows across our customer projects. We've chosen to build this framework internally to maintain maximum control over the information, services, and LLMs that power our agents. This approach also allows us to deploy our solution across various hosting services and infrastructure, tailored to meet our customers' specific needs and requirements.

In this article, we'll explore our decision-making process, explain what AI agents are, and share our current progress in this field.

Disclaimer

This article discusses our internal development of AI agents and frameworks. While we share our approach and considerations, we recommend reviewing any AI implementation with your technical and legal teams to ensure it aligns with your organization's needs and compliance requirements.

Development Status

This framework is currently under active development. The approaches and implementations discussed may evolve as we continue to refine our solution.

Security & Privacy

All implementations of AI agents should be carefully considered in terms of data privacy and security. We strongly recommend reviewing any AI implementation with your security and compliance teams.

What are AI Agents to us

We view AI Agents as temporary, 'on-demand' team members that assist in performing automated reasoning tasks with limited scope and knowledge. While we maintain traditional validation approaches for ensuring data meets specific requirements and thresholds, we employ AI agents to analyze perform tasks on the content itself.

These tasks are carefully scoped and monitored to prevent sensitive information leakage and to manage the hallucinations and overreach that AI systems typically exhibit.

Not every task requires an AI agent, so when creating workflows, we maintain control over when and how AI should be used to improve organizational processes.

Sensitive Information Workflows

AI models still train on and retain information from prompts, meaning they can and do learn from the data you send them. What you send to AI models is crucial and must align with privacy laws and obligations.

Let's explore an example workflow we've created at Goldmund, simulating a health insurance company's claim processing system.

Traditional Workflow:

User submits a request with manual information
User uploads photos of receipts
Insurance worker reviews the manual information
Worker enters information from the receipts
Internal system and guidelines validate the claim
Decision is made and communicated to the user

Enhanced AI-Assisted Workflow:

User submits request with manual information
User uploads photos of receipts
System (non-AI) anonymizes the claim information
AI analyzes claim content against user coverage for initial assessment
System (non-AI) anonymizes receipt data, extracting only billing information
AI processes the sanitized receipt data and compares it with coverage
AI populates the system with extracted information
Insurance worker reviews AI summary and receipts for final assessment
Decision is made and communicated to the user

While this workflow has more steps, much of it is automated both with and without AI. The insurance worker can focus more on claim substance rather than administration. We carefully control what information reaches the AI, keeping it limited and scoped to ensure no personal details are exposed. The worker maintains final decision-making authority while AI assists with non-sensitive tasks and preliminary assessments.

Complex Iterative Workflows

Beyond linear workflows, we've developed more complex systems where AI can make suggestions, improvements, or iterate over workflow aspects.

TV & Network Broadcaster Workflow:

User submits a TV proposal
System (non-AI) validates information completeness
AI 'Review' Agent analyzes the proposal:
- Checks for existing similar shows
- Identifies potential copyright issues
- Compares with similar shows and their ratings
- Validates categorization
- Provides improvement suggestions
If needed, 'Iteration Agent' refines the proposal:
- Adjusts titles and character names
- Updates categories
- Addresses copyright concerns
- Cycles through review process until criteria are met
Original and revised proposals are presented to the user and broadcaster
Broadcaster reviews with enhanced insights for faster, more confident decisions

This workflow employs multiple AI agents that review each other's work, providing broadcasters with comprehensive insights from the start.

A Tool, Not a Replacement

At this stage, AI Agents can only reliably function as tools to enhance existing workflows. They're not yet capable of fully replacing human oversight. While we've experimented with full automation, concerns about data security, information handling, and AI's tendency to hallucinate make complex workflows unreliable without human supervision.

Our framework integrates with third-party APIs, supports multiple LLMs and models, enables collaborative iterative workflows, and prioritizes user information security. But these agents only still help to improvem the workflow, they don't replace the need for human oversight and validation.

Security Concerns

AI agents are vulnerable to exploits from user-submitted data or exposed instructions. We've implemented several safeguards:

AI agents operate in the background behind protected APIs
Agents are instructed to never execute input data directly
Each agent is limited to specific tasks and tools
Agents are isolated to prevent cross-contamination
Access to information is strictly controlled

This compartmentalization prevents agents from accessing or influencing other agents in the workflow, maintaining strict control over their capabilities and scope.

Conclusion

Looking Forward

We continue to develop and refine our framework, expanding integrations and improving our tooling. As we progress, we'll share more insights and updates on our approach to AI agents.

If you're interested in learning more about Goldmund's work with AI agents or exploring potential collaborations, we'd love to hear from you.

Want to reach out?

Did you find the article interesting, want to discuss it or ways that Goldmund might be able to help you? Please feel free to reach out using one of the methods below.