How to Connect Business Data to an AI Agent

Start with a user job

“Connect our Drive” is not a success criterion. “Find the current refund policy and cite the approved document” is. Write ten representative questions, the source each should use, who may ask it and what a safe failure looks like. These become the initial evaluation set.

Choose sources based on those jobs. One valuable corpus with clear ownership is a better first release than ten shallow connectors.

Build the knowledge path

1. Ingest: preserve connector and external IDs.
2. Normalize: extract usable content without erasing provenance.
3. Segment: form retrieval units suited to the document.
4. Index: support semantic, exact and metadata constraints.
5. Retrieve: return a compact evidence set.
6. Deliver: expose it through REST or MCP.

The hybrid-search guide explains why semantic and exact retrieval work together.

Make permissions part of retrieval

Filter candidates to what the caller may access before evidence reaches the model. Tenant isolation in storage is not enough if the search index can surface another tenant’s chunks. Carry workspace, source and document policy into the searchable representation.

Authentication identifies a caller. Authorization determines allowed actions. Retrieval enforcement determines which evidence enters context. Treat them as separate testable controls.

Preserve and test evidence

Each passage should carry a stable document ID, title, connector identity, relevance signal and useful location. Then require answers to cite only returned evidence. A plausible answer without traceable evidence should fail.

Use the RAG scorecard to measure retrieval relevance, coverage and faithfulness separately.