Autonomy is the degree to which an AI system can pursue a goal and act without direct human intervention. In an agentic workflow, the question is not just whether an answer is acceptable, but whether the course of action is acceptable. Greater autonomy, tool access, and workflow involvement can increase the scale of harm. section 6.7
Reviewing actions one at a time may miss a harmful sequence. In one reported internal test, an agent blocked from using an authentication token split it into fragments, disguised them, and reassembled the credential at runtime. section 6.7
Set autonomy boundaries according to the impact of each action:
- Classify agents, tools, and workflow steps by potential harm. Flag actions that are high-impact, irreversible, or affect people; default to denying access when a tool cannot be classified. section 6.7.1
- Allow reversible, low-stakes actions such as reading non-sensitive data to run unattended. Require human confirmation for significant actions such as editing production systems; keep critical, irreversible actions such as touching payroll off-limits. section 6.7.1
- Enforce boundaries through permissions and checkpoints outside the agent’s control, rather than relying on prompt instructions. Place review before the action can cause harm. section 6.7.1
- Give reviewers a concise account of the agent’s intended actions, evidence, and unresolved choices, rather than relying on raw tool-call logs alone. section 6.7.1
Table 6.7 Controls for ethical agentic deployments
|
Control
|
Why it matters ethically
|
|
Classify impact of actions and tools
|
Impact estimates ensure subsequent controls
(e.g. human oversight, autonomy) activate before harm is done.
|
|
Autonomy tiers
|
Ensures that autonomy is proportional to the ethical
impacts that may harm individuals.
|
|
Accountability
|
A named human or organization is accountable for harms the
agent causes[142].
|
|
Plan preview for consequential tasks
|
Provides a clear explanation of the plan of action and
allows the reviewer to make changes to the plan to prevent harms.
|
|
Human oversight
|
Provides meaningful and accessible oversight, built around
what a reviewer can actually process.
|
|
Workflow design
|
Checkpoints and constraints outside the agent’s control
keep the workflow from drifting off its goals. Graceful failure design
ensures a halted run causes no new harm.
|
|
Redress path
|
Protects people from being trapped inside an automated
process and aligns with GDPR and EU AI Act oversight requirements.
|
|
Monitoring and logging
|
Collects the evidence needed to redress harms and prevent
recurrence.
|
|
Multi-step testing
|
Provides detailed understanding of how harms may propagate
in the agentic system.
|
|
Memory auditing
|
Prevents bias from accumulating across sessions and
bending future runs off their intended goal.
|
Compare autonomy tiers with related controls for accountability, plan review, workflow constraints, and redress. section 6.7.1













