AI agents need a shared language for business metrics

AI agents rarely fail in enterprise analytics because they cannot produce an SQL query. The larger risk is that they answer a question with a metric defined differently by finance, sales, and operations. Recent MCP and agent updates therefore make a previously underestimated architecture component visible: a shared language for business metrics.

An agent needs a data product, not a list of tables

Google Cloud’s current documentation for the Looker-managed MCP server describes an MCP integration embedded directly in Looker. This lets an MCP client access Looker data models and Looker-governed access. The protocol name is not the key issue. The key issue is that the agent receives a curated business interface instead of navigating freely across tables, field names, and duplicate dashboards.

Before an agent answers “How is our margin developing?”, the organization must define which margin applies, which period, currency, entity, and cancellation logic are included, and which role may see which level of detail. A technically correct query can still yield a business-wrong answer without those decisions.

Context is also an operational parameter

In a current article on Claude Code sessions, Anthropic explains that model choice, tool definitions, project instructions, read files, and command output enter the conversation context. Changes to the model or effort level can interrupt the prompt cache; unnecessary context therefore affects both answer quality and consumption.

The observation extends beyond coding tools. An analytics agent should not receive every dashboard, data dictionary, and tool description for every question. It needs the smallest context that makes an answer explainable: an approved metric, permitted dimensions, data freshness, authorization, and the required output format.

What DACH organizations should define now

Start with three concrete artifacts:

Metric contract: definition, owner, calculation logic, freshness limit, and allowed dimensions for every KPI.
Agent interface: only the MCP tools and data models required by the use case; no generic read permission for the warehouse.
Answer evidence: every management answer states data freshness, the definition applied, and exceptions.

This is not an extra documentation layer around a chatbot. It is the condition for an answer to retain the same meaning in steering, audit, and month-end close. Organizations that create the metric contract before a pilot can use agents for analysis without spreading inconsistent terms faster.