From '90 Days to Shutdown' to Most Efficient Unit in the Company: An Agentic AI Case Study

I was sitting in a conference room on the fourteenth floor when the CFO said it. Not with malice — with math. “This unit has ninety days to hit profitability targets or we fold it into shared services.” That was a Tuesday morning in late January. By the following Tuesday, I was on-site with their operations lead, pulling apart workflows like an engine on a shop floor.

The unit was a 34-person claims processing team at a mid-market insurance carrier. They handled commercial property claims — the kind involving adjusters, contractors, policyholders, and about forty emails per claim before anyone gets paid. The team was processing 112 claims per week with a 14.3% error rate. Their cost-per-claim was $487, nearly double the industry benchmark.

What Was Actually Broken

The team was spending 61% of their time on work that wasn’t claims processing. They were chasing documents. They were reformatting data between systems that didn’t talk to each other. They were writing the same seven emails over and over.

The people were fine. The workflow was a graveyard of manual handoffs that had accumulated over eight years of patching without redesigning.

This wasn’t a case for a chatbot. This was a case for agentic AI — systems that could reason, coordinate, and act within defined boundaries.

Three Agents, One Orchestrated Workflow

We deployed a multi-agent orchestration system built on Claude, designed around three specialized agents.

Agent One: Intake & Document Intelligence. This agent handled every inbound submission. It parsed documents, extracted key fields, cross-referenced them against the policy database, and flagged gaps. What took a human analyst twenty-plus minutes happened in under ninety seconds.

Agent Two: Compliance & Routing. This agent ran claims against regulatory requirements, internal underwriting guidelines, and historical patterns. It assigned a risk score, then routed the claim to the appropriate human reviewer.

Agent Three: Communication & Coordination. This agent managed ongoing dialogue with policyholders, adjusters, and contractors. It drafted status updates, responded to routine inquiries, and escalated anything requiring human judgment.

The three agents communicated through a shared orchestration layer. Human reviewers stayed in the loop at every decision point that carried material risk.

The Numbers

At the sixty-day mark:

- Claims processed per week: 112 → 274 (145% increase)

- First-pass documentation error rate: 14.3% → 2.1%

- Average cost-per-claim: $487 → $213

- Average cycle time: 11.2 days → 4.7 days

- Customer satisfaction score: up 31 points

The unit didn’t just hit its Q3 targets. It hit them six weeks early.

The Part Nobody Expected

Not a single person was let go. With agents handling high-volume coordination work, the team’s role shifted. Senior analysts spent time on complex claims requiring genuine expertise. The team lead started building a knowledge base that fed back into the agents’ reasoning.

The team went from cost center under threat to the most efficient unit in the company.

Most companies don’t have a technology problem. They have a deployment problem. They have the tools. They don’t have the map.