📄 SSRN Preprint — June 2026

Bridging Conventional Data Warehouse and BI Infrastructure to AI-Ready Environments via the Model Context Protocol

A Reference Architecture for Governed, Agent-Consumable Data Platforms

Diogo de Lima Souza IEEE Senior Member Senior Data Engineer & Data Architect Independent Researcher DataCompound

Abstract

Enterprises operating mature data warehouse and business intelligence (DW/BI) stacks face a structural barrier when adopting large language model (LLM)-powered analytics: their data surfaces, optimised for SQL clients and BI tools, lack the governed, tool-addressable interface required by AI agents. This paper proposes the MCP Data Bridge Architecture (MCPDB), a four-layer reference architecture that retrofits an MCP control plane onto existing DW/BI infrastructure — specifically Azure SQL Database, Azure Data Lake Storage Gen2, and Databricks — enabling any MCP-compatible AI agent to securely consume, query, and act on enterprise data without dismantling incumbent investments.

17–23pp improvement in LLM query accuracy with semantic layer grounding

O(N+M) integration surface vs. O(N×M) with ad hoc connectors

100% governance fidelity across all 5 evaluated scenarios

67–69% projected query accuracy vs. 45–51% with raw schema only

The DW/BI–AI Readiness Gap

A 2025 Cloudera survey found that 96% of enterprises had integrated AI into core processes, yet a majority reported that transformative value remained elusive due to limitations of their existing data infrastructure. MuleSoft's 2025 connectivity benchmark reinforces this: 95% of IT leaders say integration issues directly impede their AI adoption plans.

The root cause is architectural. Data warehouses optimise for columnar storage, SQL access, and human-readable dashboards. They do not natively expose the tool-addressable, governed, schema-rich surfaces that LLM agents require. Three specific failure modes emerge:

Semantic Ambiguity Column names like arr_usd_contracted or churn_flag_30d encode institutional knowledge obvious to a trained analyst but opaque to a language model. Empirical benchmarks demonstrate a 17–23 percentage point accuracy gap between LLMs given a raw schema versus a curated semantic context document.
The N×M Integration Problem Connecting M agent systems to N data tools requires up to N×M custom connectors. Each integration is point-to-point, brittle, and difficult to govern. MCP resolves this structurally, reducing O(N×M) to O(N+M).
Governance Bypass Ad hoc agent integrations bypass the RBAC, dynamic data masking, and audit logging mechanisms built into the warehouse, creating a governance shadow: AI agents acting on data with no audit trail, no masking enforcement, and no access logging.

The MCPDB Reference Architecture

MCPDB is a four-layer reference architecture organised around a governed MCP control plane. The MCP boundary sits between the semantic/governance layer and the agent/consumption layer. Existing BI tools retain their SQL/DAX connections throughout — MCPDB is additive, not a replacement.

L4 · Agent & Consumption Layer

LangGraph · CrewAI · PydanticAI · Databricks AutoML · Power BI · Looker · Claude / GPT-4o · Custom agent workflows

MCP Control Plane — Governed Tool Boundary

Tool discovery · OAuth token propagation · RBAC enforcement · DDM pass-through · Audit logging · HITL gates · Prompt-injection defence

L3 · Semantic & Governance Layer

dbt Semantic Layer (MetricFlow) · Unity Catalog Metrics · Azure Purview / Microsoft Fabric · OpenMetadata · Data contracts · Vector store (AI Search)

L2 · Transformation & Orchestration Layer

dbt Core / dbt Cloud · Azure Data Factory · Databricks Workflows · Kafka + Spark Structured Streaming · Databricks AI Functions in SQL

L1 · Storage & Table Format Layer

Azure Data Lake Storage Gen2 · Delta Lake (Unity Catalog) · Azure SQL Database · Databricks Lakehouse · Apache Iceberg

The MCP Server as Governance Enforcer

The MCP server is an active governance layer, not a pass-through proxy. It forwards the agent's OAuth 2.0 token to the warehouse on every call (RBAC applies), enforces dynamic data masking, logs every tool call to a Unity Catalog audit table, and gates write-capable tools behind human-in-the-loop approval flows.

Semantic Layer as Agent Grounding

Certified metric definitions (dbt MetricFlow YAML) are served to agents as MCP Resources before query formulation. This ensures agents always use the approved calculation and filter logic rather than inferring them from raw schemas — the mechanism responsible for the 17–23pp accuracy improvement.

How MCPDB Compares

MCPDB evaluated against three alternative migration strategies:

Strategy	Governance	Accuracy Gain	Migration Cost	BI Continuity
Full re-platforming	High	+17–23pp	Very High	Disrupted
Direct API exposure	Low	+4–9pp	Medium	Preserved
LLM-native warehouse	Medium	+15–20pp	High	Partial
MCPDB (proposed)	High	+17–23pp	Low–Medium	Fully preserved

Practical Guidance for Data Architects

MCPDB suggests a concrete four-step additive sequence:

1. Establish a governed unified catalog Unity Catalog or Azure Purview. Register all warehouse assets before exposing anything via MCP.
2. Build the dbt Semantic Layer Cover the top business metrics. This is the highest-leverage AI-readiness investment you can make today.
3. Deploy the MCP server Read-only tools available to authorised agents; write-capable tools HITL-gated. BI tools are untouched.
4. Implement the first agentic use case Pipeline health monitoring or metric anomaly investigation: well-bounded scope, immediate value, low risk.

Keywords

Model Context Protocol Data Warehouse Modernisation AI-Ready Data Infrastructure Semantic Layer LLM Agents Databricks dbt Delta Lake Azure Agentic Analytics Governance Unity Catalog

Read the Full Paper

Published as an SSRN preprint, June 2026. Full text including all tables, reference implementation details, and 23-entry bibliography.

Read on SSRN Discuss this work