Start With the Work, Not the Threat List

A list of risks is less useful if it never names the actual job. Follow a task from the employee’s question to the saved result. An AI assistant may read a document, summarize it, call another tool, and write a customer record. Each crossing changes who can see or change information.

Pick a task people already want to do, not a hypothetical super-agent. Write down the owner, the approved record to check, the least information needed, the tool’s permissions, and the action a person must approve. A narrow map reveals more than an organization-wide ban that employees cannot apply to Tuesday’s work.

Confidentiality means the wrong person sees data. Integrity means a record or answer is changed incorrectly. Availability means a process stops when people need it. A renewal assistant can fail in all three ways: expose a private note, update the wrong account, or block the team on a deadline.

Rank those harms for this task. A wrong internal summary may be recoverable; a sent price commitment may not be. This ranking tells you where to put a review, which access to remove, and what must be logged. It also prevents security from treating every draft as if it were a wire transfer.

Revisit the map when a connector—the link that lets one application read from or act in another—changes, or when the model, vendor plan, or action permission changes. A read-only pilot is not the same system after someone enables email sending. Keep the current version with the team’s workflow documentation and give the change an owner.

This course adds ten controls to that map. Lesson nine turns those same areas into a scored checklist. You do not need to buy all ten at once. You do need to know what work you are protecting, so later scores describe the real process rather than a policy binder.