Applying COMET to Business Operations: How AI Can Own Invoice Processing While Humans Keep Strategic Decisions
At ARKONA, we’re building an autonomous multi-agent AI ecosystem, initially focused on cyber-physical reverse engineering but rapidly expanding into areas like business operations, AI governance, and personal productivity. While the core principle of autonomy applies across all domains, the *how* looks different depending on the context. This post details how we’re applying our COMET framework – a 7-step human↔AI delegation methodology – specifically to invoice processing, allowing AI to handle the bulk of the work while maintaining human oversight for strategic exceptions and business rules.
The Problem with Traditional Invoice Processing
Invoice processing is a surprisingly large drain on resources for most organizations. It’s repetitive, rules-based, prone to human error, and a significant bottleneck in accounts payable. Traditional automation attempts, like OCR and basic workflow engines, often require substantial manual intervention to correct errors, handle exceptions, and validate data. This is because these systems lack the contextual understanding and reasoning capabilities to handle the inherent complexity of real-world invoices.
Introducing COMET and Autonomous Agents
COMET – our framework, grounded in IEEE and NIST standards for responsible AI delegation – provides a structured approach to transferring tasks from humans to AI agents. The 7 steps are:
- Define the Scope: Clearly delineate the task boundaries and acceptable error rates.
- Agent Selection: Choose or build the appropriate agent(s) for the task.
- Knowledge Integration: Provide the agent with necessary data and contextual information.
- Workflow Orchestration: Design the interaction between the agent and existing systems.
- Monitoring and Evaluation: Track performance metrics and identify areas for improvement.
- Exception Handling: Define rules for escalating complex cases to human reviewers.
- Continuous Improvement: Refine the agent’s capabilities and workflow based on feedback.
For invoice processing, we built a suite of agents within our BizOps domain, leveraging our inter-agent communication broker. These agents work together to form a pipeline:
- Invoice Ingestion Agent: Monitors a dedicated email inbox (currently handled via encrypted internal network) for new invoice PDFs.
- OCR & Data Extraction Agent: Uses a combination of Tesseract and a fine-tuned LLM (Ollama’s Llama3) to extract key data fields like invoice number, date, vendor, amount, and line items. This agent is a core component of our hardware RE pipeline, adapted for document processing.
- Vendor Validation Agent: Cross-references the extracted vendor information against our ERP system (currently a mock system for testing, but planned integration with NetSuite).
- Duplicate Invoice Detection Agent: Compares the current invoice details against a database of previously processed invoices to prevent duplicate payments.
- Anomaly Detection Agent: Utilizes a NIST 800-30 grounded risk evaluation engine to identify invoices that deviate from expected patterns (e.g., unusually high amounts, new vendors).
- Approval/Rejection Agent: Based on the validation and anomaly detection results, either automatically approves the invoice or flags it for human review.
- Audit Trail Agent: Records all actions taken by the agents and generates a complete audit trail for compliance purposes. All actions are cryptographically signed using SHA-256 to ensure provenance.
Technical Implementation Details
Our architecture heavily relies on MuXD, our hybrid LLM router. For invoice processing, MuXD intelligently routes tasks between local Ollama models (for speed and cost-effectiveness) and Claude (for complex reasoning and context understanding). For example, simple data extraction might be handled by a local Llama3 model, while resolving ambiguous vendor names or interpreting complex invoice terms falls to Claude. This token savings optimization is crucial, given the volume of documents we process.
Communication between these agents is handled via our message broker. Here's an example of a message published by the OCR & Data Extraction Agent:
{
"message_type": "invoice_data",
"invoice_number": "INV-2026-04-01",
"invoice_date": "2026-04-01",
"vendor_name": "Acme Corp",
"total_amount": 123.45,
"line_items": [
{"description": "Widget A", "quantity": 1, "unit_price": 100.00},
{"description": "Shipping", "quantity": 1, "unit_price": 23.45}
],
"source_file": "invoice.pdf",
"timestamp": "2026-04-07T10:00:00Z"
}
This message is then consumed by the Vendor Validation Agent, Duplicate Invoice Detection Agent, and Anomaly Detection Agent. Each agent performs its specific task and publishes updated messages with their findings. The entire process is orchestrated by a workflow engine built using our internal task delegation framework. The workflow engine exposes a dedicated API and administration interface for orchestration.
Exception Handling and Human Oversight
The critical aspect of COMET is knowing when to hand control back to a human. Our exception handling rules are designed to flag invoices that:
- Fail vendor validation.
- Exceed a predefined anomaly threshold (based on amount, vendor, or other criteria).
- Contain ambiguous or unclear information that the AI cannot confidently interpret.
Flagged invoices are routed to a dedicated queue for human review. Our 5-agent newsroom editorial pipeline’s fact-checking component is adapted to handle these exceptions, providing a second layer of verification. Humans review the flagged invoices, correct any errors, and approve or reject them. The human decisions are then fed back into the system to improve the AI’s accuracy and reduce the number of future exceptions. This closed-loop learning process is key to achieving true autonomy.
Current Status and Results (as of 2026-04-07)
Currently, 21 out of 22 services within ARKONA are online. We’ve seen 186 commits in the last 7 days, highlighting the rapid pace of development. With our invoice processing agents, we’ve achieved approximately 85% automation, reducing manual processing time by over 70%. The remaining 15% requires human intervention, primarily for complex or unusual invoices. We’re continuously monitoring the system’s performance and refining the agents’ capabilities based on real-world data.
Challenges and Future Work
One of the biggest challenges we faced was ensuring data quality and consistency. Invoices come in various formats and often contain errors or inconsistencies. We addressed this by implementing robust data validation rules and utilizing advanced OCR techniques. Another challenge was handling ambiguous language and contextual understanding. This required fine-tuning our LLMs and providing them with extensive training data.
Future work includes:
- Integrating with additional ERP systems.
- Expanding the scope of automation to include purchase order matching.
- Developing more sophisticated anomaly detection algorithms.
- Incorporating machine learning models to predict future invoice volumes and optimize resource allocation.
Key Takeaway
Applying COMET isn't about replacing humans entirely. It’s about strategically delegating repetitive tasks to AI agents, freeing up humans to focus on higher-value activities like strategic decision-making, exception handling, and process improvement. The 7-step framework forces you to think critically about the boundaries of automation and the importance of human oversight. Successfully implementing this requires a solid AI infrastructure, robust inter-agent communication, and a commitment to continuous learning and adaptation. It’s a challenging endeavor, but the potential rewards – increased efficiency, reduced costs, and improved compliance – are significant.