What Enterprise Leaders Should Require Before Giving AI Agents Access to Company Data

An AI agent that can query company data is part of decision infrastructure. The question is not only whether it can write SQL, but under what conditions its answer deserves trust.

1. A known identity

Every request should be attributable to a human or service identity. The default is simple: no identity, no schema and no answer.

2. Permission before retrieval—and execution

Apply authorization before schema reaches the model and again against every object and operation in the parsed query. Use source-native least privilege as an independent boundary.

3. Certified business meaning

Material terms such as revenue, active customer and qualified pipeline require versioned definitions with owners, grain, period, currency and exclusions. Ask when meaning is unresolved.

4. A deterministic query decision layer

The model proposes a query; it should not decide whether the query may run. Parse statement structure, constrain actions, resolve identifiers, authorize objects, compile where practical and enforce execution limits.

5. Verification proportional to consequence

An exploratory count and a number entering a forecast do not deserve the same verification. Report what actually occurred, and never turn verifier unavailability into a pass.

6. Refusal as a designed outcome

Test answerable, definition-dependent and refusal-required questions separately. A safe refusal neither invents evidence nor confirms restricted objects.

7. Evidence that survives the conversation

For material answers, retain the question, identity, policy decision, SQL, objects, semantic version, data state and verification result—with access controls on the record itself.

A public implementation example

Mnemiq, developed by Agentic Fabriq, is useful because its core paths are publicly inspectable. Open code makes claims testable, but it does not certify an unknown deployment. Identity, source permissions, definitions and operations remain customer responsibilities.

This briefing is part of the independently maintained Enterprise Data Agent Governance project. Its use of Mnemiq as an implementation example does not imply a formal partnership or independent product validation.