Every AI Agent Demo Stops at Email. I Pointed Mine at the Bills That Cost You Money.
Understanding AI Agents: From Email to High-Stakes Tasks
The Common Starting Point of AI Agents
- Many AI agent demonstrations begin with managing email and calendar tasks, addressing a common pain point for users.
- Users often find themselves stuck after initial setup, unsure how to transition from basic task management to more complex responsibilities like insurance and healthcare.
Building an Agent Skeleton
- The session aims to create an agent skeleton that can handle both low-stakes (email/calendar) and high-stakes (insurance/tax) tasks effectively.
- Problems are categorized by domain (health vs. taxes), but fundamentally they require similar organizational skills from the agent's perspective.
Key Principles in Agent Functionality
- Effective organization is crucial for extracting structured insights from unstructured data, which is a common issue in high-trust paperwork scenarios.
- Focus on agents that alleviate bureaucratic burdens rather than just performing simple actions; preparation is key.
Core Functions of the Agent Skeleton
Essential Features of the Agent
- The skeleton will perform nine functions: context packing, ingestion, chunking, normalizing, storing, retrieving, citing, exporting, and gating.
- A critical aspect is ensuring the agent does not submit or sign documents autonomously; this responsibility remains with the user.
Structure of Builds
- Three builds will be demonstrated: starting with email/calendar tasks before progressing to insurance appeals and tax preparations as advanced use cases.
Build One: Managing Email Effectively
Challenges in Email Management
- Users often struggle with disorganized inboxes filled with various types of correspondence that need structuring.
- Important documents like W2 forms or denial letters may be buried within chaotic email threads.
Context Pack Creation
- The agent creates a context pack defining what it can read from an email thread while preparing a reply for scheduling meetings.
Importance of Trust in AI Handling
Building Trust Through Transparency
- After drafting replies or proposals, the agent provides receipts detailing its sources and changes made—this transparency fosters trust between users and AI systems.
Transitioning to Higher-Stakes Tasks
Moving Beyond Basic Agents
- Transitioning from simple agents to handling complex tasks doesn't require starting over if built correctly; existing structures can be leveraged across different domains.
Build Two: Handling Insurance Appeals
Structuring Complex Information
- In this build, the focus shifts to creating detailed case files rather than vague appeal letters by breaking down denial letters into manageable components.
Normalization Process
- Dates become dates,
- Missing documents are identified,
- All information is stored locally for easy access without relying on external models.
Evidence-Based Approach
Validating Claims
- The system produces timelines and evidence checklists that help validate claims against policy language cited by insurers.
Build Three: Tax Preparation
Efficient Tax Document Management
- This build utilizes previously gathered data from emails to prepare reviewable packets for tax filing instead of directly submitting returns.
Preparing for CPA Review
- The output includes income summaries and expense ledgers along with questions tailored for CPAs—emphasizing better inquiry over mere answers.
Conclusion: Building a Scalable System
Key Takeaways Across Builds
- Clean normalized data underpins all builds; when structured properly, less expensive models can perform effectively without needing complex solutions.
- Emphasis on understanding where human expertise is necessary—especially in sensitive areas involving money or health—and avoiding one-off solutions by building scalable systems.