Contextual Agentic

Who Is Accountable When an AI Agent Acts?

Five decisions leaders can make before delegating consequential work to an AI agent.

·5 min read

Written as at 30 October 2026.

agentic-aigovernanceexecutive-leadershipaccountability

C-Suite Agentic Enterprise · Pilot issue · Draft for author review.

An agent recommends a supplier payment. The model’s answer may be useful, but it cannot accept responsibility for what happens next. Someone in the organisation must own the business outcome, decide the limits and know how to stop the process.

Illustrative scenario: a fictional composite written to make the point concrete. It does not describe a real organisation, and its figures are not reported outcomes.

Imagine an accounts-payable team using an agent to investigate a supplier’s request to change bank details. The agent reads the request, finds the supplier record and prepares an update. A confident explanation is not sufficient evidence that the request is genuine, or that the agent has permission to update it.

Five decisions belong to leadership

1. Name the outcome owner. A business owner remains accountable for the supplier-payment process, including exceptions. Technology, security and operations have their own responsibilities, but a list of teams is not a substitute for an owner who can resolve conflicts.

2. Set the agent’s boundaries. Decide which records it may read, which actions it may recommend, and which it may execute. For this example, the agent might prepare a change but be prohibited from changing bank details directly. Limits should be specific enough to test, and reviewed when the workflow changes.

3. Define approval and stop authority. Identify the role that can approve an update, what evidence it must see, and who can pause the workflow. An approval for one supplier and proposed account should not silently become permission for a different update. A human review step needs a workable escalation path, not just a button.

4. Ask for evidence of the decision and outcome. The record should make it possible to inspect the request, the relevant evidence, the authority check, the approval and the actual system response. Avoid collecting unnecessary sensitive information; determine retention and access rules with the appropriate owners.

5. Prepare for failure and recovery. Decide how teams will respond if an update is wrong, an approval expires, or the system cannot confirm whether the change occurred. Test the process, including notification, containment and reconciliation. A retry should not create a second unintended effect.

What should the board or executive sponsor see?

Start with a small set of measures: actions attempted, approved, denied or escalated; exceptions awaiting resolution; time to detect and recover; and cost per accepted business outcome. These are proposed management measures, not proof of safety. Define their scope and review the evidence alongside service quality and customer impact.

The NIST AI Risk Management Framework provides a voluntary basis for structuring AI risk management. The five decisions above are a Contextual Agentic interpretation for this example, not a NIST-mandated checklist or a guarantee of compliance.

A question to take into the next review

For every agent that can affect a customer, supplier or employee, can we name the outcome owner and show what happens when the proposed action is denied or goes wrong?

The answer is more useful than a count of agents deployed. Delegation can speed up work, but accountability and the ability to intervene remain organisational choices.

This article is part of CAE Field Notes, a series accompanying the book, The Contextual Agentic Enterprise.

References

  1. NIST, AI Risk Management Framework 1.0. https://www.nist.gov/itl/ai-risk-management-framework

← All articles

Related