Practical controls.
THE ESSENTIALS
Approved tools, named Individual accounts, not shared Strong authentication Access removed when people leave A written statement of what may not be sent
WHY BANNING FAILS
Staff use these tools anyway, on personal accounts, with no oversight and no record.
A permissive policy with clear limits produces better control than a prohibition.
WHAT TO MONITOR
Not conversations, which is disproportionate.
Whether the policy is understood, and whether output is being verified.
FOR INTEGRATIONS
Any AI feature inside software you use may process data.
Check what is enabled, particularly features enabled by default without announcement.
FOR AGENTS AND AUTOMATION
Grant the minimum capability required.
Require approval for anything irreversible.
THE PROMPT INJECTION RISK
Systems processing external content may encounter instructions hidden in it.
Constrain what automated systems can act upon.
WHAT TO REVIEW
Annually, and whenever tools change.
