Skip to main content
The risk profile changes when AI moves from generating text to taking actions in real systems.

Start with the data

Classify what the system can access:
  • public information;
  • internal business information;
  • personal data;
  • commercially sensitive material;
  • credentials or secrets;
  • regulated or high-impact information.
Collect and expose only what the use case requires.

Review tool permissions

For an agent or automated workflow, list every action it can perform. Ask:
  • Can it read only, or write?
  • Can it send messages externally?
  • Can it create, modify or delete records?
  • Can it initiate payments or commitments?
  • Can it access multiple clients or workspaces?
  • Can permissions be scoped to the minimum needed?

Separate reversible from irreversible actions

A draft that a person reviews is different from an autonomous action that changes a production system. Increase controls as reversibility decreases.

Consider prompt injection and untrusted content

If the system reads email, websites, uploaded files or third-party content, treat that content as potentially untrusted. Do not allow external text to silently override system rules or grant new permissions.

Use approval boundaries

Examples:
  • draft but do not send;
  • recommend but do not approve;
  • create a proposed CRM change but require confirmation;
  • allow low-risk routine actions only within explicit thresholds.

Log important actions

Where the use case warrants it, retain enough evidence to understand:
  • what action occurred;
  • which data or instruction caused it;
  • who approved it;
  • whether the action was changed or reversed.

Vendor and integration review

Check how authentication, data retention, model providers, third-party tools and logging actually work. Do not assume a familiar vendor name removes the need for architecture-level review.

Escalation

High-impact use cases involving employment, finance, health, safety, legal rights or regulated decisions require stronger specialist review. This guide is operational guidance, not legal or security certification.