Skip to main content
Glama

Scaling Enterprise GenAI with MCP (Bloomberg Case Study)

Written by on .

Agentic Ai
mcp
GenAI
Bloomberg
  1. Pragmatic Scaling of Enterprise GenAI with MCP
    1. Problem
      1. Intuition
        1. Architecture Example (TypeScript-inspired)
          1. Key Takeaways
            1. Acknowledgements
              1. References

                Pragmatic Scaling of Enterprise GenAI with MCP

                In this session, Sambhav Kothari from Bloomberg shares how their internal GenAI platform evolved to production-grade scale using the Model Context Protocol (MCP). This journey highlights how standardization and infrastructure principles accelerated GenAI deployment across 9,500+ engineers1.

                Problem

                When GenAI went mainstream in 2022, Bloomberg teams could quickly build demos—but productionizing those same applications was painfully slow. Delays stemmed from:

                • Too many handoffs across engineering, legal, product, and compliance
                • Custom integrations for every application
                • Lack of standardized interfaces for agents calling tools and services

                The result was a significant "productionization gap" that limited AI velocity2.

                Intuition

                To solve this, Bloomberg applied dependency inversion using MCP. Their thesis:

                If agents, tools, and LLMs adhere to standardized, configurable interfaces, they can be swapped and scaled freely3.

                They treated prompts and toolchains as config, built remote multi-tenant MCP servers with identity-awareness, and separated agent logic from app logic. This allowed them to focus on tool development and enable broad reuse across departments.

                Image

                Architecture Example (TypeScript-inspired)

                import { createServer } from 'mcp-server'; import { authMiddleware } from './infra/mcp-middleware'; import { tradeSummaryTool } from './tools/tradeSummary'; import { complianceCheckTool } from './tools/complianceCheck'; createServer({ tools: [tradeSummaryTool, complianceCheckTool], middleware: [authMiddleware], model: 'bloomberg-llm-v3', io: 'http', });
                export const authMiddleware = async (ctx, next) => { if (!ctx.headers["x-user-id"]) { throw new Error("Missing identity header"); } ctx.identity = ctx.headers["x-user-id"]; return next(); };

                Key Takeaways

                • MCP enables dependency inversion at org scale
                • Tool modularity + identity-aware middleware = secure, scalable deployment
                • Separating roles unlocks contributions from thousands of engineers
                • A configuration-first mindset closed the demo-to-production gap

                Bloomberg reduced experimentation cycles from weeks to minutes and democratized GenAI development14.

                Acknowledgements

                This guide is based on Sambhav Kothari’s talk at “Pragmatic Scaling of Enterprise GenAI with MCP” at the MCP Dev Summit 1. Special thanks to the Anthropic team and the broader MCP developer community for building production-ready agent frameworks.

                References

                Footnotes

                1. [Session] Pragmatic Scaling of Enterprise GenAI with MCP with Sambhav Kothari at Bloomberg ↩ ↩2 ↩3

                2. MCP Dev Summit – Pragmatic Scaling of Enterprise GenAI with MCP (Slides) ↩

                3. Closing the Agentic AI Productionization Gap: Bloomberg Embraces MCP ↩

                4. Enterprise GenAI Meets Real-World Scale — LinkedIn Post by MCP Dev Summit ↩

                Written by Om-Shree-0709 (@Om-Shree-0709)