Human-led
AI supplies information; a person decides and acts.
Control floor: Source quality, disclosure and normal professional review.
Decision rights and controls
Govern the action, not only the model.
An agent can turn information into consequential action at machine speed. Governance must define who authorized that action, where it must stop, and how the enterprise can intervene.
Governance principles
Authority should be explicit, bounded and revocable.
Good governance is not an approval committee. It is an operating capability that keeps action inside an authorized envelope.
01 · Named accountability
Every agent and delegated decision has a human owner accountable for its operating result.
02 · Bounded authority
Scope is defined in operational terms, tied to conditions and revocable without negotiation.
03 · Least privilege
Identity and permissions match the task, remain visible, and expire when no longer needed.
04 · Usable evidence
Material actions leave a record that leaders, operators and assurance teams can reconstruct.
05 · Meaningful intervention
High-consequence work has tested pause, override, rollback and redress—not a theoretical kill switch.
06 · Fresh authorization
New capabilities, tools or contexts do not silently inherit approval from an older system.
Authority ladder
Increase autonomy only as evidence and control mature.
Autonomy is not a property of a model. It is permission granted for a defined decision, workflow and operating condition.
Recommend
AI proposes an action; a named person approves each material step.
Control floor: Evidence threshold, approval record and clear rejection path.
Act within bounds
An agent executes reversible actions inside an explicit operating envelope.
Control floor: Least privilege, limits, monitoring, escalation and tested rollback.
Operate by exception
Agents coordinate routine work; people manage exceptions and policy.
Control floor: Continuous assurance, outcome monitoring and independent override.
- Purpose: the enterprise outcome this authority serves.
- Owner: the named human accountable for operating performance.
- Scope: the decisions and actions permitted in operational terms.
- Context: the information it may use and its provenance requirements.
- Limits: thresholds, prohibited actions and time-bound permissions.
- Escalation: conditions that return work to a person or safer mode.
- Evidence: how action, outcome, intervention and change can be reconstructed.
If an authority record cannot be explained to an operator, auditor and board member, it is not yet a usable control.
Board reporting
Report changed authority—not adoption theatre.
Show material changes in delegated authority, exceptions, control failures, unresolved exposure, intervention performance and lessons applied. Prompt counts and model usage do not answer whether the enterprise remains in control.