Governed agentic AI for back-office document processing

Context

An insurance subsidiary within a larger financial group processed over 15,000 documents per month: claim declarations, supporting evidence, medical reports, policy amendments, regulatory correspondence, and contract renewals.

The processing was almost entirely manual. A team of 35 back-office operators opened each document, identified its type, extracted relevant information, checked it against policy rules, and entered it into the management system. Peak periods (storm damage, regulatory deadlines) created backlogs of 3-4 weeks, directly impacting customer satisfaction and regulatory compliance.

The group had invested in a traditional OCR and rules-based automation platform two years earlier. It handled 15% of documents - the simple, standardized ones. Everything else fell back to manual processing because the static pipeline could not handle ambiguity, varying formats, or documents requiring cross-referencing against policy data.

Leadership wanted AI-powered automation but had clear constraints: full traceability for regulatory audits, human validation on high-stakes decisions, and zero tolerance for incorrect claim approvals.

The problem

The challenge was not document reading - modern LLMs handle that well. The challenge was building a system that could:

• Orchestrate multi-step processing: a claim document requires reading, classification, information extraction, cross-referencing with policy data, rule application, and routing. Each step has different accuracy requirements
• Handle ambiguity: real documents are messy: handwritten notes, partial information, contradictory statements, attachments in multiple formats
• Maintain governance: every decision that affects a policyholder must be traceable, auditable, and reversible
• Integrate with existing systems: the management system, the document archive, the policy database, and the compliance reporting tools were all legacy systems with limited APIs
• Keep humans in the loop: not as a fallback, but as a deliberate architectural choice for high-impact decisions

A simple "chatbot that reads documents" would not survive the first audit. The organization needed an agentic system - multiple specialized agents coordinating to process documents end-to-end - with the governance rigor of a banking-grade system.

What CMX delivered

Phase 1 - Process analysis and agent architecture (4 weeks)

We analyzed the entire document processing workflow by observing operators, mapping decision trees, and categorizing documents by complexity and risk level.

We designed a multi-agent architecture with four specialized agents:

• Classifier Agent: receives incoming documents, identifies type (claim, amendment, correspondence, regulatory), extracts metadata, and routes to the appropriate processing pipeline
• Extraction Agent: reads document content, extracts structured information (dates, amounts, policy numbers, medical codes, damage descriptions), and flags ambiguities for human review
• Validation Agent: cross-references extracted data against policy rules, coverage limits, regulatory requirements, and historical patterns. Produces a compliance assessment with confidence scores
• Routing Agent: based on document type, complexity score, and confidence levels, routes the document to: automatic processing (high confidence, low risk), human review queue (medium confidence or high risk), or exception handling (anomalies, potential fraud indicators)

Each agent had defined boundaries, explicit input/output contracts, and guardrails preventing autonomous action on high-stakes decisions.

This architecture drew on CMX's research into thought-action spaces and multi-layered cognitive systems. Each agent operates at a specific level of abstraction - the Classifier at the strategic level (what kind of problem is this?), the Extraction and Validation agents at the tactical level (what specific steps do I need?), and the Routing agent orchestrating across levels with bidirectional feedback.

Phase 2 - Agent development and testing (8 weeks)

We built each agent using a structured approach:

Model selection and optimization:

• Evaluated multiple LLMs against 500+ real documents (anonymized) across all document types
• Selected a combination: a fast model for classification and routing, a more capable model for extraction and validation
• Built structured prompts with document-type-specific instructions, validated against ground-truth datasets

Orchestration layer:

• Built the agent coordination system: sequential processing pipeline with parallel branches for independent checks
• Implemented state management: every document's processing state is persisted, enabling pause, resume, and rollback
• Defined escalation rules: when any agent's confidence drops below threshold, the document exits automated processing and enters the human review queue with full context

Guardrails and governance:

• Output validation rules on every agent: format checks, value range checks, consistency checks between agents
• Complete audit trail: every agent decision, every tool call, every confidence score logged and queryable
• Human override capability at every stage: an operator can intervene, correct, and re-route any document at any point
• Automatic fraud pattern detection: anomalous patterns flagged for specialized review

Phase 3 - Integration and deployment (6 weeks)

We integrated the agentic system into the existing operational environment:

• Document ingestion: connected to the existing document capture system (email, postal scanning, web uploads) via event-driven integration
• Management system integration: API integration with the legacy policy management system for data retrieval and update. Built adapter layer to handle the system's SOAP-based interface
• Human review interface: built a review dashboard where operators see the document, the agent's extraction and assessment, and can validate, correct, or override with one click
• Reporting and compliance: automated generation of processing reports, audit logs, and regulatory compliance documentation

We deployed in progressive waves: first one document type (simple amendments), then expanded to claims, then to the full document portfolio. Each wave included a parallel-run period where both manual and automated processing ran side by side for validation.

Phase 4 - Optimization and handover (2 weeks)

After production stabilization:

• Tuned confidence thresholds based on real production data to optimize the automation-vs-review balance
• Built a feedback loop: operator corrections feed back into prompt refinement and validation rule updates
• Trained the internal team on agent monitoring, prompt management, and model evaluation
• Delivered operational documentation and incident response procedures

Results

• 70% of documents processed automatically end-to-end without human intervention (up from 15% with the previous OCR system)
• Average handling time reduced by 65%: documents that previously took 25 minutes average now take under 9 minutes (including those routed to human review)
• Error rate below manual baseline: the agentic system's extraction accuracy exceeded human operators on standardized documents, and the human-in-the-loop design caught edge cases
• 4 governed agents in production with full audit trails, meeting regulatory audit requirements
• Peak period backlogs eliminated: the system scales with volume, processing surge periods without degradation
• Operator role transformed: from data entry to quality assurance and exception handling, improving job satisfaction and retention

Why this matters

Agentic AI is not about replacing humans with autonomous robots. It is about designing intelligent systems where AI handles volume and pattern recognition, and humans handle judgment, exceptions, and governance.

The critical difference between a demo and a production agentic system is architecture: clear agent boundaries, explicit guardrails, traceable decisions, human escalation paths, and integration with real enterprise systems.

CMX designed this system as enterprise architecture, not as an AI experiment. That is why it passed regulatory audit, why operators trust it, and why it handles 15,000 documents per month in production.