Summary of "Preparing IT for AI Agents: How MCP Shapes the Future of AI"

High-level thesis

Current enterprise IT (applications + data lakes + network, connected by APIs) is ill-suited to the new AI/LLM era. Jamming large language models into that architecture produces very high failure rates (~90%+). To succeed, IT must be re-architected to be “AI-ready” by adopting orchestration, agent-based operation, and a Model Context Protocol (MCP) approach that mirrors how the human brain integrates specialized systems.

Key technological concepts and vocabulary

AI swallows the internet vs. AI should integrate the enterprise LLMs trained on broad internet data are fundamentally different from the narrow, sensitive, operational data inside organizations. Enterprise AI needs different integration patterns and safety controls.
Brain/body-plan analogy (design metaphor)
- Lower brain: primitive sensing and fast responses
- Midbrain: connectivity, routing, memory selection
- Upper brain (frontal/executive): integration, goal setting, strategic decisions
- Important properties to emulate: selective attention (ignore most data), strong cross‑modal integration, compartmentalization, and low‑power compactness
Model Context Protocol (MCP) A proposed service interface for applications and data to expose:
- Tools: actions an application can perform
- Data sources: what the application or dataset “knows” (contextual, partitioned access)
Orchestration layer A new executive layer that spawns and coordinates many AI agents; analogous to the brain’s frontal lobe.
Agents as synapses Lightweight autonomous workers that call MCP services, access partitioned data, and perform actions across applications and data sources.
AI-ready data layer vs. unstructured data lake Partitioned, organized, contextualized data (not a “swamp”) so agents and models get clean, reliable access.

Problems with the existing approach

Point-to-point API / star architectures are brittle and require tightly structured integrations for every flow.
Jamming AI into this model yields poor outcomes and very high failure rates for enterprise initiatives.
Unorganized data lakes make agent/model access unreliable, unsafe, and context-poor.

Proposed architecture and practical implementation guidance

Follow incremental, non-disruptive adoption while introducing new layers and interfaces.

Don’t rip everything out Introduce new layers (orchestration, MCP hosts) without disrupting existing operations.
Add an orchestration layer Define goals/outcomes and spawn/coordinate agents to achieve them.
Replace the “API-as-everything” mindset with MCP-enabled services
- For each application (CRM, HRIS, finance, CLM, etc.) expose an MCP host that serves tools (actions) and data contexts.
- Partition and organize the data lake into an AI-ready data layer and make it accessible via MCP.
Design agents to be goal/outcome driven
- Agents should operate toward explicit goals and acceptable outcomes.
- Use MCP to discover and call tools and data sources across the estate.
Treat applications as specialized “organs” Keep domain-focused apps intact but make them reachable by orchestration and agents for integrated tasks.
Move from ad-hoc projects to an architecture that improves success rates The aim is a systematic platform that raises AI initiative success (target cited: 80%+).

Benefits and intended outcomes

Better cross-application and cross-data integration, modeled after human-like selective attention and executive coordination.
Scalable, flexible automation via many lightweight agents instead of brittle point-to-point APIs.
Safer, more contextual model access to organizational data and capabilities.
Higher probability of successful enterprise AI projects and more predictable orchestration of AI-driven workflows.

Tutorial / guide elements contained in the talk

A conceptual guide for re-architecting enterprise IT to support AI agents.
An actionable checklist:
- Add an orchestration layer
- Partition the data lake into an AI-ready layer
- Implement MCP hosts for applications
- Expose tools and data contexts via MCP
- Spawn and coordinate agent networks
Design rationale using the brain analogy to justify compartmentalization, orchestration, and selective data exposure.

Metrics and claims

Cited current failure rate: >90% for AI initiatives when AI is simply jammed into legacy IT.
Target success rate after adopting the proposed architecture: 80%+.