// Accountability · AI agents

Who is accountable when an AI agent acts?

The institution that delegates authority remains answerable for the decision—but accountability is credible only when it is supported by evidence and the power to intervene.

Assigning an accountable owner does not close the governance gap. The institution must also define what the agent was authorised to decide, preserve the evidence behind its actions and ensure that someone can detect, stop and correct material divergence.

// 01

The accountability illusion

Institutions often allocate ownership separately: technology owns the platform, risk owns the policy, the business owns the outcome and a vendor owns part of the model or toolchain. Every component has an owner. The eventual decision may not.

The Fiduciary Gap™: the structural gap between what an autonomous system is optimised towards and what the institution is obliged to protect. It widens when an agent’s executional power grows faster than its contextual encoding.

// 02

Four tests of real accountability

Mandate

Can the accountable person state what the agent was authorised—and forbidden—to do?

Evidence

Can they produce the data, rules, tools, judgement and authority behind one action?

Intervention

Could they have stopped or corrected the action while consequences remained alterable?

Outcome

Did feedback return to someone able to change the chain?

Join ownership

Who owns the hand-offs between policy, model, tool, agent and execution?

Replay

Can the institution demonstrate rather than merely assert control?

// 03

A delegation contract

For each consequential agentic decision, the institution should define the action space, autonomy level, decision class, escalation triggers and evidence requirements. Irreversible or conditionally reversible decisions require stronger boundaries and earlier intervention points than routine reversible actions.

The governing principle is simple: authority may be delegated; institutional accountability cannot be outsourced.

// Questions people ask

Common questions

Is the human supervisor automatically accountable?

A title alone is insufficient. Accountability requires visibility into the decision, authority to intervene and evidence of what occurred.

Can responsibility sit with the AI vendor?

Contractual responsibilities may be allocated, but the institution still has to answer for decisions executed under its mandate and affecting its customers or obligations.

What should be documented before deployment?

The agent’s mandate, action space, autonomy, escalation triggers, prohibited actions, evidence payload and intervention authority.

// The practical test

Test one institutional decision

The ten-day Decision Drift Audit™ maps one material decision across all eight layers, assesses replayability and authority boundaries, and delivers one prioritised board finding.

Test one decision →Canonical definitions →Decision Integrity Chain™ →