Least privilege
Agents should only receive the access required for their role.
AUTONOMOUS WHERE USEFUL. HUMAN-CONTROLLED WHERE IMPORTANT.
AI agents can be extremely useful.
They can also make mistakes.
The answer is not unlimited autonomy. It is clear boundaries.
HOW WE APPROACH AGENTIC SYSTEMS
Defined approval boundaries
Routine lower-risk actions may be automated. Strategic, commercial, privacy-sensitive and higher-risk actions require human approval.
Auditability
Important actions and recommendations should be recorded.
Reversibility
Changes should be versioned and reversible where practical.
Verification
Important factual or operational outputs should be checked before becoming authoritative.
Clear responsibilities
Agents should have defined roles rather than unrestricted access across an organisation.
AI IS A CAPABILITY, NOT AN EXCUSE TO REMOVE JUDGEMENT.
The purpose of agentic systems is to reduce unnecessary friction and increase organisational capability.
Human judgement remains essential where decisions affect trust, strategy, privacy, reputation or significant commercial risk.