All guides
Governance checklistAI operations

An AI governance checklist for real-asset teams

The practical controls to review before an AI workflow can read sensitive data, recommend a decision, or change a system.

9 min read
01

Define the boundary of authority

Write down the eligible users, data, tools, records, and outcomes for the workflow. A production agent should not infer its authority from a broad prompt. Authentication, tenant isolation, roles, field permissions, and approval rules must be enforced by the application and API.

02

Make every consequential result reviewable

Material recommendations need source evidence, deterministic calculations where applicable, and a clear explanation of unresolved uncertainty. Writes should use narrow validated tools, and approvals should show exactly what will change before it changes.

  • Inventory data sources and the permissions inherited from each.
  • Define prohibited data and high-consequence uses.
  • Test missing, conflicting, malicious, and out-of-scope inputs.
  • Log tool calls, approvals, corrections, and final outcomes.
  • Assign an owner for exceptions, incidents, and periodic review.
03

Govern the operating loop

Models, prompts, schemas, and source systems change. Monitor workflow-level performance, not only model metrics, and revalidate when a material component changes. A governance process is complete only when it can pause, investigate, correct, and safely resume the work.

Read next

From scattered files to a governed knowledge layer

Continue reading