Connect It to Live Data

The context layer can now explain Acme. It still needs a safe way to reach live CRM records. Keep three jobs separate: access, interpretation, and enforcement.

MCP supplies a connection contract

An MCP server can expose tools, resources, and prompts to an AI application. It can authenticate a connection and give the model a structured way to request CRM data. It does not know what “qualified pipeline” means in your company or which abandoned field looks authoritative.

Context shapes the plan before execution

For “How much qualified pipeline did we create last quarter?”, load fiscal dates, qualification criteria, exclusions, currency handling, and the allowed fields. The planner can then produce an exact date range and filters instead of translating business language on the fly.

The tool boundary checks the finished plan

Validate the final objects, fields, operation type, and user scope in code. This happens after planning and before the query runs. A model that plans an unapproved field gets a deterministic error. It does not get to substitute a similar field.

question
  → always-load context
  → relevant rule bodies
  → query plan
  → tool-boundary validation
  → CRM query
  → grounded answer + trace

Assemble that sequence in the GTM Lab. A valid tool call is not proof of a valid business answer.

Reference

• MCP vs. context layer →
• One CRM question traced end to end →