Accountability means having identifiable responsibility for decisions, actions, and their consequences. An autonomous system may make and execute decisions, but it cannot bear institutional accountability for their legitimacy or effects. The enterprise must retain a traceable chain of delegation and ownership. section 2.1.2
An account freeze at 3 a.m.
A fraud-response agent freezes customer accounts without a clear authority chain, explicit decision boundaries, or a runtime accountability trail. Even if the agent acts as designed, the enterprise must know who authorized the freeze, when escalation is required, and who is accountable for customers left stranded. section 1.1
Delegated authority can range from a recommendation requiring approval to action followed by human review. Each arrangement needs explicit boundaries and ownership; technical permission to act is not the same as authority to decide. section 1.1 section 6.7
Make the delegation chain explicit enough to answer:
- Who authorized the autonomy, and who owns the business purpose it serves? section 2.1.2
- Who can widen, narrow, or suspend its limits? section 2.1.2
- Who remains accountable for harm, disputes, or correction after deployment? section 2.1.2
In asynchronous coordination, accountability must survive the handoff. An event stream should preserve the originating interaction, the authority or constraint that shaped it, and the next accountable participant. A broker can carry signals, while durable workflow or case state records progress and an evidence store retains supporting evidence. section 6.4.2
Working integration is not proof of accountable design. A raw API exposed as an agent tool can bypass authority; an opaque event flow can move information without revealing who caused what. Governed wrappers, explicit coordination, durable state, evidence trails, and designed human reentry paths restore ownership without banning the underlying tools. section 6.7 section 6.4.2
















