Five questions anyone can ask of an AI agent.
Boards, regulators, journalists and the people affected by an agent should not need a specification to hold it to account. The 5 As are the plain-language version: if an operator cannot answer all five, the agent is not ready to act.
The five questions
Who authorised this agent, on whose behalf does it act, and within what limits?
A good answer shows: A named principal, a written mandate, stated limits and an expiry.
Maps to GUARD Grant · PACTE Purpose & Authority · AIT-12 T5
What can it actually reach — systems, data, money, people, other agents?
A good answer shows: Effective permissions reconciled against declared authority. Gaps listed.
Maps to GUARD Data · PACTE Authority & Capability · AIT-12 T5–T6
What did it do, step by step, and is there a record someone else can check?
A good answer shows: A tamper-evident log of actions, tools called and decisions taken.
Maps to GUARD Audit · PACTE Tactics · AIT-12 T4–T5
What changed in the world as a result — and can it be stopped, reversed or contained?
A good answer shows: A record of external effects, and a tested route to interrupt and revoke.
Maps to GUARD Usage & Revocation · T-RUE · AIT-12 T5–T6
Who answers for it, and how does an affected person get an explanation and redress?
A good answer shows: A named accountable person, a contact route, and an incident and redress process.
Maps to GUARD Audit & Revocation · PACTE Evidence · Incident & Disclosure protocol
The 5 As and GUARD
GUARD — Grant, Usage, Audit, Revocation, Data — is the normative control set in F-ACT 1.0, the thing an implementer declares conformance against. The 5 As do not replace it and carry no conformance level. They are the questions a non-specialist asks; GUARD, PACTE and AIT-12 are where the answers are specified, tested and evidenced.