Who is accountable when an AI agent makes a mistake?
A named person is accountable, never the agent. Usually that is the agent's owner, the person responsible for what the agent is for and what it may do. If a human approved the specific action, that approver answers for that decision too. And if nobody was ever named, you have a ghost agent, and the accountability lands on whoever deployed it.
The question usually gets asked too late, after something has gone wrong. By then the useful answer is not "who do we blame", it is "who can explain this, and who can change what happens next". That has to be set up before the mistake.
Three people who can be accountable
| Who | Accountable for | When they are the answer |
|---|---|---|
| The owner | What the agent is for, and the authority it was given | Almost always. The agent acted inside powers someone gave it |
| The approver | A specific decision they signed off | When the mistake went through a human approval |
| The deployer, by default | Everything, because nobody else was named | When the agent has no owner: a ghost agent |
The third row is the one that hurts. An agent with no owner does not make accountability disappear. It moves it to whoever pressed deploy, often someone who has since changed roles or left.
Why "the AI did it" is not an answer
An agent can tell you what it did. It cannot answer for it: it cannot explain why the authority it used was appropriate, and it cannot decide what the organization changes next. "The AI did it" describes the event. It does not name who is responsible. We wrote about this in more detail in why "an AI did it" is not an audit answer.
Set it up before the mistake
Rendering diagram…
- Name an owner. One person, not a team alias. Write down what the agent is for.
- Give it a role, not a person's account. An agent that runs on someone's personal credentials inherits everything that person can do, and becomes a ghost when they leave.
- Put a human approval at the decisions it must not make alone. Money, credentials, accepted security risk, changes of direction.
- Keep the approval decisions. Who approved or rejected what, so the approver can be found later.
A quick self-check
Run this for each agent you run today:
Agent: ______________________
Owner (a person): ______________________
Role / what it is for: ______________________
What it can touch: ______________________
Decisions that need a human, and who approves: ______________________
If any line is blank, that is where accountability will be missing on the day something goes wrong.
What this looks like in agent.ceo
In agent.ceo, an agent is a role your organization holds, one of 16 predefined roles or a custom one (see agent roles), not a person's account. When an agent files a proposal, it waits in the organization's approvals queue. An admin approves or rejects it, and that decision is kept, so the approver of a decision can be found afterwards.
The platform does not decide who the owner is. That is a choice your organization makes, and it is the one that matters most.
FAQ
Who is accountable when an AI agent makes a mistake?
A named person, never the agent. Usually that is the agent's owner: the person responsible for what the agent is for and what it is allowed to do. If a human approved the specific action, that approver shares the accountability for that decision. If no owner was ever named, the organization has a ghost agent, and the accountability falls on whoever deployed it by default.
Can you hold an AI agent itself accountable?
No. Accountability means someone can explain a decision, answer for it and change what happens next. An agent can report what it did, but it cannot be responsible for it. Saying the AI did it describes the event; it does not name who answers for it.
How do you set up accountability for AI agents before something goes wrong?
Give every agent a named owner and a defined role, limit what it can touch to what that role needs, put a human approval at the decisions it should not make alone, and keep each approval decision. Then when something goes wrong, the question of who answers for it already has an answer.
Related: What is the ghost agent problem? · Where should a human approve in an AI agent workflow?