GridArray Editorial
AI agent data readiness checklist
A practical checklist for deciding whether the data behind one agent workflow is trustworthy enough for production use.
Direct answer
Start with one workflow, not the whole estate
Agent data readiness is not a company-wide cleanup score. It is a judgment about whether the sources, permissions, retrieval path, and access contracts behind one real agent workflow are trustworthy enough to support production decisions and actions. If you cannot name the workflow, the systems it depends on, and the failure modes that would hurt the business, you are not ready to expand scope.
Framework
The agent data readiness checklist
Score each item against the specific workflow you want in production. A gap on any critical path item is a production risk, even if the rest of the estate looks polished.
1. Source authority
Identify every system or document collection the workflow must trust. For each one, name the owner, the freshness expectation, and the rule that wins when records conflict.
- Workflow source inventory exists and is current
- Authority rules are written for known conflicts
- Stale or duplicate sources are explicitly excluded
2. Structure
The agent needs records and documents in a shape it can use. Unstructured archives may be searchable, but production workflows usually need fields, entities, relationships, and stable identifiers.
- Required entities and fields are defined
- Identifiers remain stable across systems
- Free-text sources have an extraction or review path
3. Provenance
Every retrieved snippet or derived claim should point back to a source an operator can inspect. Without provenance, review becomes guesswork and incidents become debates.
- Retrieved context carries source identifiers
- Derived outputs retain a reviewable trail
- Operators can open the underlying record quickly
4. Permissions
The agent should inherit the least access required for the workflow. Broad service accounts and shared credentials turn a useful tool into an uncontrolled reader or actor.
- Identity is explicit for the agent and tools
- Permissions match workflow scope
- Sensitive actions require stronger validation or human review
5. Retrieval quality
Retrieval that returns something is not the same as retrieval that returns the right thing for the decision at hand. Evaluate the path the workflow actually uses.
- Golden questions exist for the workflow
- Missing-context failures are visible
- Retrieval changes are regression-tested before release
6. Interfaces and validation
APIs and MCP tools need contracts. Inputs should be validated before irreversible actions, and outputs should be checkable against expectations the business already understands.
- Tool and API contracts are narrow and documented
- Validation blocks unsafe or incomplete actions
- Audit logs capture who or what acted and why
Diagnosis
Symptoms versus root causes
Symptoms
- The agent contradicts itself across sessions.
- Reviewers cannot tell which source justified an answer.
- A useful demo collapses when permissions or schemas change.
Root causes
- Multiple systems of record with no authority rule for the workflow.
- Retrieval without provenance or evaluation.
- Over-broad access and weak validation on tools that can act.
Implementation
How to use the checklist in practice
Run the checklist with the people who own the workflow, the data, and the release decision. Mark each item ready, at risk, or blocked. Then sequence engineering around the blocked items that sit on the critical path to production, not around the largest enterprise backlog.
- Pick one agent workflow with a real user or business dependency.
- Complete the checklist in a working session, not as a survey.
- Separate nice-to-have data projects from production blockers.
- Connect remaining gaps to governed access, evaluation, and operations work.
Boundaries
Where this checklist stops
- It does not justify broad data lake projects without a workflow boundary.
- It does not replace evaluation, observability, or incident controls.
- It does not claim that MCP alone makes internal data production-ready.