Case study · Logistics & Mobility
AI-native analytics: grounding LLM agents in governed metrics with MCP
Connecting LLM agents to a governed metric store through the Model Context Protocol so teams can self-serve analytics conversationally.
Why this matters
Self-serve dashboards answer the questions someone predicted. Business users also ask questions nobody predicted, and those still become report requests. LLM agents can answer them conversationally, but only if their answers are grounded in the same governed definitions as the rest of the company.
The approach
The key design choice is to connect agents to the governed metric layer, not to raw tables.
- Governed metrics first. Metric definitions come from the Metric Store and Semantic Reporting Layer that already power self-serve reporting.
- MCP as the interface. The Model Context Protocol exposes governed metrics and data to LLM agents through a standard, auditable interface.
- Guardrails inherited, not reinvented. Access controls, PII masking and audit trails from the governance program apply to agent queries too.
- Capability in-house. The Analytics Engineer function was set up partly to build and maintain this AI analytics capability inside the team.
Expected outcomes
- Teams self-serve conversationally instead of raising manual report requests.
- Answers stay consistent with dashboards, because both read the same definitions.
- Analysts spend more time on decisions and less on data retrieval.
An LLM agent is only as trustworthy as the metrics it is grounded in. Governance is the foundation AI-native analytics is built on, not an afterthought.
This program is in progress. I will publish results here as they land.