A policy may say that an AI assistant can draft a response but cannot send it. The architecture review is where the team can demonstrate whether that boundary actually holds.

NIST SP 800-218A extends the Secure Software Development Framework with practices and considerations for generative AI and dual-use foundation models. It provides a development-security reference, not a guarantee that a particular deployment is safe or compliant. Read NIST SP 800-218A.

Walk through a realistic request

A useful review exercise is to trace one request from the user to the final action. Identify the documents retrieved, the services called, the credentials used, and the points where information or authority crosses a boundary.

Then ask the team to demonstrate an exception: a request for a restricted document, an unavailable dependency, or an attempted action outside the assistant’s approved purpose.

Make policy statements testable

The following are suggested review questions, not a claim that any one control is sufficient:

• Access: Can the assistant retrieve information that the requesting user cannot access?
• Actions: Which connected tools can change records, send messages, or start transactions?
• Approval: Where a person must approve an action, what prevents that step from being bypassed?
• Evidence: What records help a reviewer reconstruct a disputed action?
• Recovery: How can the team disable a capability and continue the business process?

An illustrative example

Consider a support assistant intended only to suggest replies. A review could ask the team to show that its service account lacks permission to send messages. If the business later requests automatic sending, that becomes a distinct design decision with new evidence to assess.

This example illustrates the review approach; it does not describe a client engagement or tested product.

Leave with decisions and evidence

The review output should state the approved purpose, unresolved design questions, demonstrations completed, and conditions for the next change. That makes the architecture discussion useful to both technical teams and the leaders accountable for deployment.

Discuss your AI governance questions with StrategyAI Partners.

Back to Insights