Understand the idea, test it actively, then explain it in your own words.
You will leave with
Evaluate what an agent may access and change separate read, create, modify, send, delete, purchase, and delegate permissions match approval and observation to consequence, scope, uncertainty, and reversibility
Time
30 min
Before you begin
General AI Literacy unit 09 is useful preparation, but the path begins with a complete system-classification foundation.
Do this now
Read the outcomes and useful terms before the first section.
separate read, create, modify, send, delete, purchase, and delegate permissions
match approval and observation to consequence, scope, uncertainty, and reversibility
Terms you will use
These definitions prepare you for the reading; you do not need to memorize them.Least privilege
Giving a system only the access and actions needed for the current purpose, for the shortest practical scope and time.
Approval gate
A point where execution pauses until an authorized person or system accepts a clearly presented action and its consequences.
Reversibility
The degree to which an action can be reliably undone, contained, compensated for, or restored within an acceptable time and cost.
01
Access is not one permission
Reading a calendar, creating a draft event, inviting participants, deleting events, and granting another account access are different powers. “Calendar access” hides this distinction. A useful permission map names object, operation, scope, duration, identity, and purpose.
Permissions should narrow with context. An assistant preparing one meeting may need read access to selected availability and permission to create a draft, not every historical event, private note, or the authority to send invitations.
02
Consequences determine where control belongs
Low-impact reversible preparation can often proceed automatically. External communication, financial commitment, access changes, publication, deletion, and decisions affecting rights usually need stronger confirmation, separation of duties, or human review.
Approval must be informed and timely. A person cannot meaningfully review a vague button after the action or inspect hundreds of hidden changes. Present the exact proposed effect, important inputs, uncertainty, alternatives, and whether reversal is actually available.
03
Control is a system, not a single button
Before action, use identity checks, least privilege, constraints, previews, and approvals. During action, apply budgets, rate limits, monitoring, and pause controls. After action, preserve suitable logs, notifications, reconciliation, revocation, and recovery.
Human involvement is not automatically effective. Reviewers need time, authority, relevant evidence, and an escalation route. When consequences are difficult to reverse, prevention and small scoped trials matter more than relying on an undo label.
DuringLimit time, volume, cost, recipients, and rate; monitor and preserve a usable stop path.
AfterConfirm actual effects, notify owners, reconcile records, and revoke temporary authority.
If wrongContain, reverse or compensate, preserve evidence, escalate, and learn without hiding the incident.
Misconception to correct
“A human-in-the-loop checkbox makes consequential agent action safe.”
Review works only when the right person sees the real proposed effect before execution, has evidence and time, and can refuse or escalate.
Active checks
Decide, then compare the reasoning
No grade, score, or streak. Progress records completed learning objects, not your worth or ability.01
An agent is helping schedule one meeting. Which permission is proportionate?
02
Which approval is most meaningful before a €2,000 purchase?
Explain back and transfer
Create a permission and control map
Choose one agentic use and list every read, create, modify, communicate, delete, spend, or delegate action. For each, define scope, duration, approval, observation, and recovery.
Answer both checks before completing.
Source context
What supports this unit and where it stops
Evidence status: established
AI Risk Management FrameworkU.S. National Institute of Standards and Technology
Supports contextual risk mapping, governance, measurement, management, monitoring, and accountability; it does not provide a complete access-control implementation.
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence ProfileU.S. National Institute of Standards and Technology
Adds generative-AI risk actions and lifecycle considerations; controls remain context-specific and the profile is voluntary guidance, not a legal guarantee.
inAi