Loading...

Media is loading
 

Paper Type

Complete

Abstract

Corporate communications teams face information overload, refresh pressure, and overreliance risk on LLM-generated briefings. We present the production deployment of an on-premises agentic LLM system for enterprise news intelligence that couples retrieval, clustering, and report generation with ensemble LLM-as-Judge quality governance. Over six months in a conglomerate (Sep 2025–Feb 2026), Corporate Monitor ingested 36,910 unique articles and the Top News pipeline executed 126 runs (52,462 articles). Reports are scored on source faithfulness, factuality, informativeness, and coherence; grade B (≥75) defines the automation–oversight boundary. Sub-threshold outputs trigger up to three retries; remaining cases are flagged for review. Results show 100% success for scheduled pipelines, 99%+ reliability across user-initiated workflows, and an 85.3% pass rate across 2,592 quality-gated outputs. A pilot expert study (40 evaluations) shows directional agreement with expert judgments and flags corporate relevance as a missing dimension. We present design principles for responsible agentic AI governance at scale.

Paper Number

1279

Comments

AI SYSTEM

Share

COinS
 
Aug 15th, 12:00 AM

Deploying Agentic LLM Pipelines at Scale: Quality-Gated Ensemble Governance for Enterprise News Intelligence

Corporate communications teams face information overload, refresh pressure, and overreliance risk on LLM-generated briefings. We present the production deployment of an on-premises agentic LLM system for enterprise news intelligence that couples retrieval, clustering, and report generation with ensemble LLM-as-Judge quality governance. Over six months in a conglomerate (Sep 2025–Feb 2026), Corporate Monitor ingested 36,910 unique articles and the Top News pipeline executed 126 runs (52,462 articles). Reports are scored on source faithfulness, factuality, informativeness, and coherence; grade B (≥75) defines the automation–oversight boundary. Sub-threshold outputs trigger up to three retries; remaining cases are flagged for review. Results show 100% success for scheduled pipelines, 99%+ reliability across user-initiated workflows, and an 85.3% pass rate across 2,592 quality-gated outputs. A pilot expert study (40 evaluations) shows directional agreement with expert judgments and flags corporate relevance as a missing dimension. We present design principles for responsible agentic AI governance at scale.

When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.