THE SITUATION
Why this decision is reaching leadership now
An AI agent can do more than produce text. Depending on its design, it may retrieve information, call software tools, change records, communicate externally or coordinate a sequence of actions.
That capability can reduce manual effort, but it also changes the governance question from ‘Is the answer accurate?’ to ‘What authority has the system been given, and what happens when its plan is wrong?’
THE IMPLICATION
What could happen if the issue remains unresolved?
Excessive permissions
Broad access can turn one incorrect instruction or malicious input into a consequential action.
Prompt injection
Content retrieved from documents, websites or messages may attempt to redirect the agent or obtain information.
Nominal human oversight
An approval step provides little protection if the reviewer lacks time, context, authority or a clear way to challenge the recommendation.
THE DECISION
What a controlled approach requires
Boundaries
Define permitted tools, data, actions, limits and prohibited outcomes.
Approval
Require meaningful human authorisation before material, external or irreversible actions.
Evidence
Retain an audit trail of inputs, sources, decisions, tool calls, changes and approvals.
Recovery
Provide safe failure, escalation, rollback and incident-response routes.
QUESTIONS TO TAKE INTO THE ROOM
How confidently could your organisation answer these?
- What precise outcome requires agentic behaviour rather than a controlled workflow?
- Which systems, records and actions can the agent access?
- Can untrusted content influence the agent’s instructions or tool use?
- Which actions require informed human approval?
- Can every consequential action be reconstructed and explained?
- How will success, failure and exception paths be tested?
- How can the organisation pause, reverse or safely retire the agent?
FROM QUESTION TO EVIDENCE
An informed question is useful. An organisation-specific assessment is more valuable.
ARI helps leadership establish whether the organisation has the governance, security, information and operating capability to move beyond experimentation responsibly.