Ready to start your AI journey? Book a Call
Back to All Case Studies
Agentic AI

Arca Logistics deployed a 12-agent system that reduced dispatch errors by 89%

Manual dispatch coordination across 400 routes was creating costly errors. We built and deployed a multi-agent system that handles scheduling, exception management, and escalation.
89%
Fewer dispatch errors
3.2x
Route efficiency gain
$5.8M
Annual value created

The Problem

Northfield Bank's credit team was spending 60% of their working day on documentation. A senior credit analyst earning $180K was spending 4.5 hours a day writing and formatting credit memos - work that required judgment at the beginning and end, but largely involved pulling information from structured loan applications, financial statements, and internal credit policies.

Three previous attempts to automate this with rules-based tools had failed because credit memos require contextual interpretation of client-specific financials against internal and regulatory standards. The bank needed something that understood their specific credit framework, not just generic document automation.

Our Approach

We started with a two-week diagnostic phase: interviewing 12 senior analysts to understand the information sources, decision logic, and quality standards that define a good credit memo at Northfield. This produced a structured spec that became the evaluation framework for everything built afterward.

The technical solution was a RAG architecture built on Anthropic's Claude, with a proprietary knowledge base comprising 12 years of Northfield credit memos, regulatory guidelines (Basel III, SR 11-7), and internal credit policy documentation. The system retrieves relevant precedents and policy context for each new loan, generates a structured draft memo, and routes it through a compliance review step before presenting it to the analyst for final judgment.

We ran a 6-week controlled pilot with 20 analysts before full deployment. Output quality was evaluated against a panel of senior analysts blind to whether each memo was human-written or AI-assisted. By week 4, AI-assisted memos were rated higher on completeness and regulatory citation accuracy.

The system doesn't replace analyst judgment. It eliminates the 3 hours of mechanical work that surrounds it. Our best analysts are now doing analysis, not formatting.

The Result

Full deployment to 140 analysts in week 11. Average credit memo time dropped from 3.8 hours to under 1 hour. The bank's credit committee noted a measurable improvement in memo quality and consistency across the team. Total annual cost savings: $2.1M in analyst time, with an additional $400K identified in error-related rework reduction.

The system now handles 200-300 memos per week with zero production incidents in the first 6 months. A quarterly model review process keeps the knowledge base current with regulatory updates.

Getting Started

Ready to build AI that actually ships?

Book a 30-minute discovery call. We will assess your highest-value AI opportunity and tell you exactly what it takes to build it.
Love the background dot animation? Explore the playground
Launch Plan Builder
Need help customizing?