How Marsh Built LenAI, Their Company Brain, at Enterprise Scale

Marsh is a Fortune 500 global company that owns multibillion-dollar businesses including Marsh, Guy Carpenter (now Marsh Re), Mercer, and Oliver Wyman. Inside such a big org, they built something unusual: a central innovation team reporting directly to the board, led by developers who could move at startup speed inside an enterprise company.

Each Marsh business unit was building AI tools in isolation while employees used personal ChatGPT accounts, creating security nightmares. The innovation team needed to unify AI development without forcing standardization that would slow everyone down.

A centralized AI innovation business unit built by the company's CIO/COO, Paul Beswick, and led by Niall Maher, built the foundation—LenAI, powered by Mastra—then enabled each business unit to build on top of it.

Why Mastra

Marsh's technical background made Mastra a natural fit. Niall Maher had joined the company to lead its Inner Source initiative (open source for enterprise), and the innovation team operated primarily as a Node.js web development shop. Even Beswick remained a contributor to its TypeScript and Rust codebases.

The recommendation came from experienced DevRel Jesse Hall, then at MongoDB and now at LiveKit. So when Maher met Hall for beers, Hall told him, "You should try Mastra. I think you'd like it," Maher immediately acted on the recommendation.

"We're throwing anything and everything into Mastra. It's made our web devs deadly at developing AI applications. It's by far the best developer experience for any of the products we've used, even in the Python space." — Niall Maher, Innovation Engineering Leader

From a chatbot to a multi-agent system

LenAI is Marsh's internal ChatGPT-like interface. It connects to all of Marsh's systems—internal databases, external APIs, web search, and other tools available within the company. Marsh was evolving it from a knowledge and chat interface into an agent that could reason about requests and orchestrate standard MCP tools through Mastra. Mastra is the connection layer between models, agents, workflows, and tools.

The innovation team has already built a bunch of specialized agents across the organization:

• Career consultant agent for personalized employee guidance
• Survey visualization agent that automatically selects the best chart type based on question phrasing
• Database explorer agent that navigates complex Databricks warehouse structures

They also have a shared approach to agent development across the org now. A sister team created Mastra templates to give developers a common starting point that says "this is how we do agents, full stop." Business units can use those templates to build their own apps and MCP tools, and LenAI organizes those capabilities behind specialist agents.

Organizing tools behind specialist agents

LenAI has a main conversational agent that orchestrates Marsh's standard MCP tools. As the range of possible tasks grew, the team began turning larger capabilities into agents of their own. A PowerPoint agent, for example, could have separate tools for generating a presentation and editing individual slides.

In general, the architecture becomes more complex as more Marsh teams add their own MCP tools. The more tools the agent could choose from, the less accurate it was. As a solution, the innovation team started grouping related tools behind specialist agents.

"We convert them to agents but make them available as tools to the big-boy agent," they say. This gives the main LenAI agent a way to hand off specialized work without carrying every underlying tool itself. The pattern they want to get to is selecting an agent that has access to a subset of tools, rather than selecting a tool.

Marsh also explored Mastra's agent-network approach, where a routing agent can choose among other agents, workflows, and tools based on the request.

What's next for LenAI

Niall Maher's team is working on various improvements to the platform. On the evals front, they want to move beyond user feedback alone and introduce formal evaluations. By storing real conversations, the team can measure whether agents select the right tools, return relevant answers, and help users complete their intended tasks.

The team is already thinking beyond that design: if LenAI eventually has a hundred specialist agents, choosing the right agent creates the same routing problem they faced with tools.

Finally, they hope to keep growing the adoption. The Mastra-powered platform was deployed to roughly 80,000 people in 3 months.

"If I walked around our office right now, I bet five or ten of the 40 people would have Mastra docs open," says Maher. Soon, he expects to see 40 to 50 Mastra-powered applications across the org.

"We've even been telling all our vendors to use it, if you're wondering why there's a spike of traffic coming out of Dublin and New York" adds Niall Maher.