Key pannellist included: 

• Mehdi Goodarzi, SVP, Global Head of Generative AI Practice - Hexaware Technologies
• Craig Brown, Director of Banking Technology - European Bank for Reconstruction and Development, EBRD

Mehdi Goodarzi emphasised the limitations of current AI models, which often provide outdated references and lack internal context, leading to misfires in organisational applications. He highlighted the shift from experimentation to implementation, illuminating the complexities of integrating AI across siloed systems and the necessity of custom coding for system connectivity. MCP, developed by Antropic and other AI leaders, promises to address these issues by standardising the integration process, providing a protocol layer that informs AI models of user needs and trusted sources, akin to a universal connector for AI tools and data sources.

Goodarzi elaborated on how MCP can revolutionise AI integration by enabling real-time access to external tools, APIs, and documents, thereby enhancing context awareness and dynamic tool discovery. He underscored the differences between MCP and other emerging protocols like A2A, which focuses on agent-to-agent communication, while MCP concentrates on connecting agents to external data sources and systems. The potential applications of MCP were illustrated through examples like financial risk assessment, where AI assistants can query diverse agents to gather comprehensive contextual data from enterprise systems, facilitating informed decision-making. Despite its promises, MCP faces challenges such as prompt injection vulnerabilities, lack of formal certification, and potential vendor lock-in, which Goodarzi suggested could be mitigated through early adopter influence and open-source engagement.

Craig Brown of EBRD provided a practical perspective on MCP's implementation. He described how EBRD's legacy systems and unstructured data necessitated a shift towards AI, leading to the creation of an AI engineering squad and the adoption of MCP. Brown likened MCP to the Corba protocol, highlighting its ability to decouple data sources from AI models, thus lowering entry barriers and fostering experimentation. He envisioned MCP enabling fully agentic software development teams, although cultural and data-related challenges persist. Brown stressed the importance of collaborative problem-solving to overcome scepticism and blockers, positioning MCP as a catalyst for innovation and empowerment within organisations.

Takeaways

AI factories require extensive infrastructure akin to small cities

AI factories need significant investments in power, cooling, and physical infrastructure, similar to planning a small city. This holistic approach involves integrating new technologies like liquid cooling and ensuring robust power supply to support massive data processing needs.

MCP as a universal connector for AI integration

The shift from experimentation to AI implementation has revealed complexities such as siloed systems and custom coding requirements. MCP addresses these issues by providing a standardised protocol that facilitates seamless connectivity between AI agents and external data sources, thereby improving decision-making.

MCP's potential in empowering AI-driven innovation

Craig Brown illustrated how MCP could lower entry barriers and foster experimentation within EBRD. By decoupling data sources from AI models, MCP creates opportunities for innovative applications like agentic software development teams, despite facing cultural and data-related challenges.