Six domains for a pre-approval review
A pre-approval review works best when every use case is examined against the same domains. The table below lists the six domains, and the sections that follow turn each one into questions the business should answer.
| Domain | What the review covers |
|---|---|
| Business purpose and data | What the system is for, who owns it, and what data it reads, writes, or sends out. |
| Identity and permissions | Which identity the agent uses and what that identity is allowed to do. |
| Tool and integration access | Which tools, APIs, and MCP servers the agent can invoke. |
| Model and vendor dependencies | Which third parties handle data, and under what terms. |
| Runtime controls and monitoring | Which constraints are enforced while the system operates. |
| Auditability and regulatory alignment | What is logged, who is accountable, and which obligations apply. |
Start with purpose, ownership, and data exposure
Begin with what the system is for and who answers for it. An AI use case without a clear purpose and a named owner cannot be assessed, because every later question depends on knowing what the system is meant to do.
Questions to ask
- What is the business purpose of the system, and who owns it?
- What data does the system read?
- What data does it write?
- What data does it send outside the organization?
Evidence to expect
- A named business owner for the use case.
- A description of the data the system reads, writes, and sends out.
Ask about agent identity, permissions, and tool access
Once the data is understood, examine what the agent can actually do. Identity determines what the agent is allowed to do, and tool access determines what it can reach.
Identity and permissions
- Which identity does the agent use?
- What is that identity allowed to do?
Tool and integration access
- Which tools can the agent invoke?
- Which APIs can it call?
- Which MCP servers can it reach?
Evidence to expect
- The identity the agent uses, with the permissions granted to it.
- A list of the tools, APIs, and MCP servers the agent can invoke.
Examine model and vendor dependencies
An AI use case rarely runs on a single system. Identify every third party involved and understand how each one handles data.
Questions to ask
- Which third parties are involved in the use case?
- Which of them handle data?
- Under what terms do they handle it?
Evidence to expect
- A list of the model and vendor dependencies.
- The terms under which each third party handles data.
Separate runtime enforcement from design-time documentation
Documentation describes how a system is intended to behave. Runtime controls determine how it behaves while it operates. A review should distinguish between the two and ask which constraints are enforced, not only which are written down.
Questions to ask
- Which constraints are enforced while the system operates?
- Which constraints exist only as documentation or intent?
- How is the system monitored while it runs?
Evidence to expect
- The constraints enforced at runtime, identified separately from design-time documentation.
- A description of how the system is monitored in operation.
Confirm auditability, accountability, and regulatory fit
Approval should not end with the decision itself. The business should be able to show afterward what the system did, who is accountable for it, and which obligations apply.
Questions to ask
- What is logged?
- Who is accountable for the system?
- Which obligations apply to the use case?
Evidence to expect
- A description of what is logged for audit.
- A named accountable owner.
- The obligations the business has identified as applicable.
Answers that should change the approval decision
Some gaps should prevent a straightforward approval. Each of the following should map to a condition, an added control, or a rejection.
- The business cannot say what the system is for or who owns it.
- The business cannot describe what data the system reads, writes, or sends out.
- The identity the agent uses, or what that identity is allowed to do, is unclear.
- The tools, APIs, or MCP servers the agent can invoke are not known.
- Third parties handle data, but the terms are unknown.
- Constraints exist only in documentation and are not enforced at runtime.
- Logging, accountability, or applicable obligations cannot be shown.