Authority before
autonomy.
Identity. Evidence. Escalation. Action boundaries. Auditability.
When an AI system can do more than answer — when it can trigger workflows, change state or initiate action — governance must reach the point where intelligence becomes consequence.
An agent needs more than
a good prompt
A capable agent may retrieve data, call tools and propose or execute a sequence of steps. The governance question is not only whether the output is useful. It is whether the agent is operating with the right evidence, within the right authority, against the right checks, and with a record that makes the action reconstructable.
Recommendation, drafting, escalation, execution and irreversible external action are different authority levels. They should not be collapsed into one permission.
Before an agent
changes anything
Identity & purpose
Which agent is acting, for which task, under which scope and on whose behalf?
Evidence requirements
What information must be present, current and verified before the action may proceed?
Counter-evidence
Which observation should block, downgrade, reverse or escalate the proposed action?
Human authority
Which actions require an identified accountable person to accept, reject or stop them?
Action boundary
What may the agent do automatically, what is read-only, and what requires explicit approval?
Shadow first.
Autonomy by evidence.
A practical first step is often read-only or shadow mode. The agent observes the workflow, produces recommendations and records the evidence behind them without being allowed to create consequential external effects.
Authority can expand only when the organisation has defined what good looks like, what failure looks like, what must be logged, and which actions remain outside the automated boundary.
This is not a claim that every workflow should become autonomous. The purpose of governance is to make the permitted boundary explicit.
If an agent acted,
you should be able to explain why
Evidence
What the agent knew and where it came from.
State
Relevant model, agent, prompt, policy or tool configuration.
Checks
Premises verified, failure conditions evaluated and counter-evidence considered.
Authority
Whether action was pre-authorised or explicitly accepted by a human authority.
Action
What changed, where, when and through which tool or system.
Outcome
What happened afterwards and what should be learned from it.
Agentic AI governance
in plain language
What is agentic AI governance?
It defines and enforces how an AI agent may observe, recommend, escalate and act, including evidence requirements, authority, action boundaries and the record needed to reconstruct what happened.
When should an AI agent require human approval?
When an action exceeds its pre-authorised boundary, creates a consequential external effect, or requires accountable judgement that has not been delegated.
What should be recorded when an AI agent acts?
Relevant evidence, agent or model state, the proposed action, checks performed, counter-evidence, authority exercised, downstream action and observable outcome.
The control layer
around the decision
Choose one agentic workflow.
Define the boundary before the action.
We can map the evidence, authority, failure conditions, permitted tools and decision record around one consequential workflow without replacing the existing stack.