Who has authority
when AI can act?
Approval. Escalation. Stop authority. Pre-authorised boundaries.
An AI system may generate a recommendation, prepare an action or trigger a workflow. That does not by itself give the system institutional authority to commit a consequential action.
Capability is not
decision authority
The operational question is not only what the AI can do. It is what the organisation has authorised it to do, under which conditions, and who remains accountable when the consequence matters.
A model may classify, rank or recommend. An agent may draft, retrieve, route, call a tool or propose a transaction. Authority begins where a proposal becomes an organisational commitment or external effect.
The boundary should be explicit enough that a reviewer can tell whether the system stayed inside it and, if it crossed the line, who had authority to approve that step.
Four states around
a consequential action
AI proposes
The system produces a recommendation, classification or candidate action. No external commitment is implied.
AI prepares
The system may draft or stage the action while keeping execution outside its authority.
Authority is requested
When the action exceeds the pre-authorised boundary, the system routes it to the identified person or role that may decide.
Action occurs within authority
Execution occurs only when the action is inside an explicit pre-authorised boundary or after the required authority has been exercised.
Name the person or role
before the decision matters
Human Acceptance Authority identifies who may accept, reject, escalate or stop a consequential AI-supported action when human accountability is required.
The authority is defined by the organisation. SINAIGROUP-AI does not assign legal or regulatory authority to individuals; we make the operational boundary explicit so it can be implemented, reviewed and audited.
The record should show which authority was exercised, what evidence was available, what action was approved or rejected and what happened afterwards.
Automation can be bounded
without becoming ownerless
Not every low-risk action needs a human click. But a pre-authorised action still needs an owner, conditions and a boundary.
Scope
Which class of action is allowed without case-by-case approval.
Conditions
What evidence and system state must be true before the action is allowed.
Limits
Value, duration, destination, data access or other limits that define the boundary.
Fail-safe
What suspends execution when confidence, evidence or system health falls outside tolerance.
Rollback
How an authorised operator reverses or contains the action where reversal is technically possible.
Override
Who can stop the system and under what circumstances.
Authority should survive
the moment it was exercised
A useful decision record can preserve the evidence available at the time, model or agent state, the proposed action, relevant checks and counter-evidence, the authority exercised, the approval or rejection, any escalation, the downstream action and the observable outcome.
That makes it possible to distinguish a model output from an authorised organisational decision months later.
Authority is one part
of the decision control layer
AI Decision Assurance
Evidence, counter-evidence and authority around one consequential decision.
Open → Agentic AIAgentic AI Governance
Govern delegated action when AI can trigger workflows and tools.
Open → AuditabilityAI Audit Trail
Reconstruct evidence, system state, authority, action and outcome.
Open →Human authority
in plain language
What is Human Acceptance Authority in AI?
It identifies the person or role with authority to accept, reject, escalate or stop an AI-supported action when human accountability is required.
Does every AI action require human approval?
No. A narrow class of low-risk actions may be pre-authorised when the boundary, conditions, fail-safes, rollback path and override are explicit. Consequential actions that exceed that boundary should not be treated as automatically authorised.
What is the difference between recommendation and authority?
A model or agent can generate a recommendation or proposed action. Authority determines whether that proposal may commit the organisation to an external or consequential effect.
What should be recorded about human approval?
The record should identify the decision, evidence available, proposed action, authority exercised, approval or rejection, any escalation and the resulting action.
Map the authority around
one consequential decision
Tell us what the AI proposes, what it may do automatically, who is accountable today and what happens if the action is wrong.