Case Study
Marsh is a leading global professional services firm focused on risk, strategy, and people. Operating in over 130 countries with 85,000+ employees and tens of billions in revenue, it advises on complex risks – from cyber to climate – and provides investment guidance, with its Mercer business advising on over $15 trillion in assets.
1,000
applications
built on Java, .NET, and dozens of other stacks and database systems
93%
accuracy of issue remediation
First-time fix rate, end to end
5x
productivity gain by dev teams
78% less effort vs manual remediation
“CAST tells AI exactly what’s impacted with every change – that's what gave us the confidence to do this at scale.”
Brian Horan
Director, Application Maintenance
CAST makes technical debt remediation at scale possible for Marsh by equipping AI with deterministic architectural context.
Challenge
Marsh manages a portfolio of nearly 2,000 applications, half of which are custom-built solutions spanning dozens of different technologies accumulated over decades of growth and acquisitions.
With CAST revealing the true extent of accumulated technical debt across this vast landscape, Marsh's central maintenance team faced an unscalable mountain – manual remediation simply could not keep pace with the volume and complexity of the issues.
Solution
Marsh build an AI-powered workflow for tech debt remediation. CAST Imaging maps architecture, dependencies, and data access, feeding LLM prompts with deterministic context. Developers queue violations; AI returns fixed code.
Initially ran by a central team, Marsh is now rolling out CAST intelligence into developer IDEs and AI agents via MCP, enabling identification and remediation of critical issues as they occur in AI-powered development workflows.
Results
Marsh achieved a 93% first-time success rate for AI-generated fixes accepted into production, compared to 10–30% when using AI alone.
Marsch measured 78% productivity gains vs manual remediation – far ahead of Stanford Research benchmark of typical 0-10% gains when using AI alone for complex tasks on brownfield applications.
Additionally, structural flaws measured against ISO 5055 fell by 63% – far exceeding the CIO expectations.
